Monocular reconstruction of wind-driven vegetation is severely underconstrained: motion along the viewing direction is largely unobservable, a moving canopy offers few reliable correspondences, and nearly the entire scene is dynamic, providing little static reference. Directly-learned deformation fields in 4D Gaussian Splatting therefore optimize photometric consistency rather than recover the motion that produced it. We replace that field with a physically parameterized deformation prior: one damped harmonic oscillator per rigid part, driven by the observed wind and integrated by differentiable RK4, supervised photometrically alone. To test whether such a prior is physically grounded rather than merely well fit, we build a controlled synthetic testbed of three procedurally generated trees spanning an order of magnitude in skeleton complexity, whose per-part natural frequency follows from its own geometry and whose damping ratio is a fixed constant, both held out of training. On it, we measure held-out views, temporal extrapolation, zero-shot transfer to unseen wind speeds, and recovery of the physical parameters themselves. The prior costs appearance fidelity on in-distribution views and extrapolates markedly better outside the training window and the training wind, while parameter recovery is far weaker than it first appears: frequency recovery survives an untrained null control on only the sparsest of the three trees, and damping is not recovered at all.
Phenotyping an agricultural crop is crucial for studying its entire life cycle, as it provides vital insights to improve yield and, ultimately, food production. Doing the same for crops grown on remote sites is a challenge for the specialists who cannot be available on-site. 3D reconstruction techniques offer a promising solution to this problem by enabling crop digitization, allowing specialists to access the resulting 3D crop models from anywhere at any time. In this work, we evaluate recent 3D reconstruction pipelines for crop phenotyping. We focus on 7 mesh reconstruction pipelines and measure the fidelity and consistency of their outputs qualitatively and quantitatively. Our results suggest that the meshes produced by the GGGS, PGSR, and 2DGS are preferable to the other pipelines, owing to their quantitative metrics and visually pleasing outputs. The GGGS pipeline is better than the second-best pipeline (2DGS) by about 27\% on the radar chart with 5 dimensions, namely, User ratings, Chamfer distance, LPIPS, PSNR, and SSIM.
Herearii Metuarea · Abdoul-Djalil Ousseini-Hamza · Walter Guerra · Francesca Zuffa · Francesco Panzeri · Andrea Patocchi · Lidia Lozano · Shauny Van Hoye · François Laurens · Jeremy Labrosse · Pejman Rasti · David Rousseau
Abstract In machine learning-driven plant phenotyping, well-annotated image datasets are essential for developing robust models capable of capturing phenological variability across environments. Here, we introduce DeepPhenoTree-Apple Edition , a multi-site and multi-variety RGB image dataset dedicated to the detection of key phenological stages in apple trees. The dataset includes 48,320 time-stamped RGB images acquired across four European orchards of the Apple REFPOP consortium under biological, agronomic, and environmental variability, including differences in genotypes, orchard architectures, phenological development, temperature, and humidity conditions. From this large corpus, a carefully curated subset of 808 representative images was manually annotated. It includes 241,600 expert annotations covering developmental stages from dormant bud to fruit maturity. Images were acquired using a standardized tractor-mounted phenotyping platform equipped with active flash illumination. Active flash illumination was used to reduce illumination variability and homogenize exposure, shadows, and sunlight differences across sites. Phenological structures were annotated following the BBCH scale, with bounding boxes adapted to organ visibility and developmental stage. In addition to the dataset, we provide deep-learning baseline experiments to illustrate detection performance and detection performance across locations.
Leaf area (LA) is one of the most important morphological traits for the assessment of plant growth, biomass production and physiological performance. For forage crops, rapid, accurate and non-destructive methods of LA estimation are particularly important to allow repeated measurements throughout the growing season. The aim of this study was to compare the leaf area estimation of sainfoin (Onobrychis viciifolia Scop.) by four regression models (linear, logarithmic, polynomial and multiple linear regression) based on simple plant morphological measurements. Ninety sainfoin plants were evaluated at the flowering stage under controlled conditions. Plant height and canopy width were measured manually and actual leaf area was measured with a LI-COR LI-3100C Leaf Area Meter. Regression analyses were performed using SPSS software to construct prediction models and compare the performance of the models. Of the models tested, the multiple linear regression model had the highest prediction accuracy (R²=0.981), which was significantly better than the polynomial (R²=0.413), linear (R²=0.399) and logarithmic (R²=0.361) models. The results demonstrated that the combination of plant height and canopy width greatly improved the leaf area estimation over the use of a single predictor. The predictive performance of simple regression models was low. Multiple linear regression was a reliable, rapid and non-destructive method to estimate the leaf area of sainfoin. These results pave the way for practical morphological models that could assist plant growth monitoring and agronomic studies while limiting the number of destructive samplings.
Accurate 3D phenotyping of rice in multi-plant scenarios is important for evaluating yield-related traits and supporting high-throughput breeding. In particular, grain-level instance segmentation from point clouds provides a direct basis for quantifying panicle structure and grain number. However, this task remains challenging because densely grown rice plants often exhibit severe inter-plant occlusion, organ adhesion, and highly similar local structures, making it difficult to simultaneously capture global plant organization and fine-grained grain geometry. To address these challenges, we propose SPINet, a structure-enhanced point cloud instance segmentation network for 3D phenotypic extraction in multi-plant rice scenes. The proposed framework introduces a RiceMamba2 encoder that combines the long-range dependency modeling capability of Mamba2 with a parallel convolutional branch for sequence-neighborhood enhancement, enabling effective representation of both global spatial context and local organ-level details. In addition, a center prior-guided decoder is developed to provide explicit spatial anchors for instance queries, thereby improving the separation of adjacent and adhered rice organs in dense scenes. To further enhance training stability and convergence, a DINO-based denoising strategy is incorporated into the optimization process. Experimental results on the controlled indoor Grouped Rice Dataset (GRD) show that SPINet achieves an average precision (AP) of 65.73%, outperforming OneFormer3D by 31.38 percentage points. Furthermore, grain counting based on the predicted instances obtains a coefficient of determination ( R 2 ) of 0.8532, demonstrating the potential of SPINet for downstream yield-related phenotypic trait extraction. These results suggest that SPINet provides a promising framework for automated 3D rice phenotyping in complex multi-plant scenarios.
Introduction Accurate extraction of maize seedling phenotypes is essential for early growth assessment, yet field-grown seedlings exhibit complex backgrounds, non-rigid leaf deformation, self-occlusion, and adhered leaves that challenge three-dimensional point-cloud analysis. Methods Here, we developed a workflow integrating PMSPA-Net, an enhanced PointNet++ model, with geometric region decomposition for semantic and organ-level segmentation of field-grown maize seedlings. Multi-view RGB images of 65 complete plants were reconstructed with OpenMVS, and all data were partitioned at the whole-plant level before point-block sampling. Fifteen plants were retained as a fixed independent test set, while the remaining samples were used for training and validation across three predefined random seeds. Results Under a common protocol, PMSPA-Net outperformed PointNet, PointNet++, DGCNN, and RandLA-Net, achieving an mIoU of 0.8286 ± 0.0149, a macro-F1 of 0.8981 ± 0.0238, and a balanced accuracy of 0.9561 ± 0.0223. A four-configuration ablation study showed complementary gains from the multi-scale spatial pyramid attention module and IOA-based hyperparameter optimization. Geometric region decomposition also achieved more accurate and complete adhered-leaf separation than DBSCAN in three annotated cases. Point-cloud-derived leaf length and width agreed well with manual measurements. Discussion This workflow demonstrates the feasibility of integrated semantic and organ-level phenotyping under the evaluated field conditions; validation across additional sites, developmental stages, and acquisition conditions is required before broader generalization.
Image-based crop phenotyping benefits from image series that capture plant development together with organ-level labels. Such paired data are limited because organ annotation is expensive, and following the same plants over time requires repeated, registered imaging. Existing generative models for plants either synthesize temporal imagery without structural labels or generate labeled images without a temporal dimension. We introduce OneGrow, a latent flow-matching model that jointly models wheat images and their organ-segmentation masks over time. Images and masks share a single frozen image autoencoder. A reveal specifies which content is observed context, so the same model covers tasks such as mask-to-image synthesis, image-to-mask segmentation, and temporal forecasting. For sequences longer than the training window, a sliding-window roll-out generates each new image from the preceding frames, keeping long sequences temporally consistent. We train jointly on a large single-frame wheat dataset and a multi-year temporal dataset, using pseudo-labels from a pretrained segmentation model. We evaluate segmentation and image quality against held-out references and assess the multi-task and temporal behavior qualitatively.
This study aimed to evaluate the growth responses of kenaf (Hibiscus cannabinus L.) under nitrogen and compost treatments and to develop a UAV-based plant height estimation model. Ground-measured plant height differences were not significant at harvest (110 days after sowing, DAS), highlighting the need for multitemporal monitoring. Multispectral drone imagery was acquired at five growth stages (20–110 DAS). Object-based image segmentation was applied to extract pure vegetation areas, and digital surface model differencing (ΔDSMt) was used to reduce micro-topographic effects. A UAV-based multiple linear regression (UAV-MLR) model was developed using ΔDSMt, NDVI, GNDVI, and NGRDI to integrate complementary structural and spectral information. Evaluated on the calibration dataset, the UAV-MLR model demonstrated high fitting performance (adjusted R2 = 0.9858, RMSE = 14.95 cm, MAE = 11.31 cm), outperforming the ground-based simple linear regression (G-SLR) model based on stem diameter (adjusted R2 = 0.9615, RMSE = 39.68 cm, MAE = 29.16 cm). By integrating structural and spectral information, the proposed approach reduced RMSE by 62.3%, offering a highly accurate, non-destructive tool for crop monitoring and precision agriculture.
Abstract Developmental progression in plants is inherently a time-to-event process, yet plant phenology data are often analyzed as ordinary endpoint traits even when some individuals fail to reach the target stage within the observation window. Such observations are right-censored rather than missing and should be retained in inference. Here, we evaluated 21 soybean cultivars across 10 controlled-environment treatment combinations varying in CO 2 concentration, photoperiod, and LED light quality, and analyzed days from sowing to the R6 stage using a practical workflow that combined Kaplan-Meier survival analysis with a Bayesian Weibull accelerated failure time (AFT) model. Based on 210 cultivar × treatment observations derived from 1,050 plants, treatments resolved into favorable, intermediate, and strongly inhibitory classes. Elevated CO 2 showed the strongest association with reproductive timing: increasing CO 2 from ambient to 1000–1400 ppm reduced median survival time, while the additional gain from 1000 ppm to 1400 ppm was minimal, indicating a saturating response. Blue - Red spectral treatments and 6–8 h photoperiods were associated with rapid and synchronized development within the tested treatment combinations. The Bayesian AFT model included treatment as a random effect to account for the chamber-level experimental structure; the between-treatment standard deviation was estimated at 0.02 (95% credible interval: 0.00–0.06), indicating negligible chamber-to-chamber variation. The highest-ranked predicted combination (1400 ppm, blue: red = 2:1, 6 h) gave a median time to R6 of 66.90 d (mean 72.18 d). This study provides a practical and transferable workflow for censored plant phenology datasets in controlled-environment research, phenotyping, and breeding. The workflow integrates established survival analysis methods and explicitly accounts for chamber-level design, offering a framework for structured plant phenotyping experiments.
Soil contamination represents a growing environmental concern, particularly in areas such as the "Land of Fires" (Campania region, Italy). This study evaluates the potential of Zea mays L. as a bioindicator of soil contamination through a multiscale approach. In bin-like containers under natural conditions, maize plants were grown on untreated (C) and artificially contaminated (T) soil with a mixture of heavy metals (Pb, Zn, Cr) and benzo[a]pyrene. Analyses were conducted at leaf, canopy, and field scales, integrating eco-physiological, spectral, and UAV measurements. At the leaf level, contamination induced functional alterations in photochemical and spectral traits that were strongly dependent on the growth stage, being most pronounced early in the season and progressively attenuating thereafter. At the canopy level, contaminated plants developed less height and leaf area, together with lower chlorophyll- and canopy-density-sensitive spectral indices, consistent with reduced vegetative vigor under contamination stress. Solar Induced Fluorescence (SIF) showed no residual difference between treatments once normalized for biomass production at the end of the growing season, in agreement with leaf-level active fluorescence measurements taken on the same date. At the field scale, UAV-based classification achieved superior performance with hyperspectral relative to multispectral data in distinguishing contaminated areas. Taken together, these results show that, under the experimental exposure conditions adopted here, the detectability of soil pollution by maize progressively shifts from early, transient physiological signals to persistent structural differences, while UAV hyperspectral data provided the highest discrimination between treatments. By tracking how leaf-level functional alterations translate into canopy- and field-scale structural separability, this study provides a scale-explicit assessment of maize as a promising bioindicator of soil contamination.
Accurate and dynamic monitoring of wheat phenotypes is essential for breeding decision-making and crop management. However, RGB image-based phenotyping still suffers from expensive pixel-level annotation, unstable organ-level segmentation across growth stages, and limited multi-trait extraction under complex field conditions. To address these issues, a high-throughput phenotyping framework (WheatScoper) was proposed, enabling organ-level segmentation and plot-level multi-trait extraction. To reduce annotation cost, a structure-aware geometry-assisted annotation (SAGA) algorithm was developed, yielding an approximately 7.6-fold improvement in annotation efficiency over fully manual annotation. To enable efficient organ-level segmentation, a lightweight semantic segmentation network (WheatScopeNet) was developed by integrating parallel hybrid spatial modeling with cross-scale feature fusion. On the held-out test set from the same site and growing season, WheatScopeNet achieved an mIoU of 0.869 and an mDice of 0.930. Leveraging the segmentation results, an automated system was established to extract 41 multi-view image-derived traits (I-traits) across four core phenotypic dimensions. The extracted I-traits supported the estimation of eight manually measured agronomic traits, with R 2 values ranging from 0.477 to 0.697. Notably, the correlations between stay-green-related-traits and yield varied distinctly with viewing-position. Only upper side-view indicators remained significantly correlated with yield, with Side-up final GPAR showing the strongest association, whereas top-view GPAR-derived indicators showed weak associations. Finally, a web-based platform integrating cascaded inference, segmentation visualization, and automatic I-trait extraction was developed. The platform provides an end-to-end solution for field wheat phenotyping and supports breeding decision-making and crop management.
Zea mays (maize) is a critical global crop, generating more than 1 billion metric tons of grain annually. Plant morphology is tightly associated with final yield and is determined by the process of development. Accurate assessment and recording of mature plant morphology is essential for linking genotype to phenotype and for assessing environmental influences on plant development and yield. In maize, imaging of fully emerged aerial organs provides an accessible and informative developmental end point for comparative phenotypic analyses. Here, we provide a protocol for capturing high-quality images of mature maize plants and detached leaf blades, suitable for publication and downstream quantitative trait extraction. We emphasize standardized imaging conditions, including high-contrast backgrounds, diffuse lighting, fixed camera positioning, and consistent scale calibration, to minimize technical variation and to facilitate automated image processing. Guidance is provided for both laboratory- and field-based imaging setups and uses widely available equipment. Images generated using this protocol can be analyzed using platforms such as FIJI to extract morphometric traits, including length, width, and outline-based shape descriptors. By prioritizing standardization at the point of acquisition, this protocol supports reproducible, scalable phenotyping of mature maize morphology for developmental studies in maize.
Reliable plant segmentation in high-throughput phenotyping must transfer across species and imaging conditions without repeated model tuning or extensive reannotation. We compare three segmentation strategies using images from Oak Ridge National Laboratory's Advanced Plant Phenotyping Laboratory: (i) fixed color-based thresholding, (ii) supervised U-Nets trained from scratch, and (iii) pretrained vision transformers fine-tuned for binary segmentation. Models were evaluated on a held-out test set and a generalization set that comprised unseen species. On the held-out test set, thresholding, the best U-Net, and the best vision transformer achieved mean Dice scores of 58.3, 96.6, and 97.3, respectively. On the generalization set, the corresponding Dice scores were 56.5, 86.2, and 95.7. Thresholding remained effective on some datasets but failed when plant appearance changed. Supervised U-Net training resolved within-distribution errors but failed to generalize to novel species and backgrounds. Pretrained vision transformers consistently produced high-accuracy segmentations across the evaluated species, views, soil backgrounds, and tray types. These results benchmark the practical progression from fixed rules to task-specific supervision and pretrained visual representations for controlled-environment plant phenotyping.
Abstract Purpose Agrivoltaic vineyards show strong spatio-temporal variability in canopy shading, but field methods to quantify panel-induced shading at canopy scale remain limited. Shading is a key factor because it affects plant physiological and morphological traits, with potential consequences for yield and production quality. This study developed a near-surface time-lapse RGB imaging approach to derive temporally explicit shading metrics in an agrivoltaic vineyard of Vitis vinifera cv. Falanghina in Southern Italy. Methods Two representative vine positions beneath the photovoltaic structure were monitored: Agrivoltaic Shade (AVS), with greater exposure to panel-induced shading, and Agrivoltaic Light (AVL), with lower exposure. Image-based canopy shading percentage was calculated through a dedicated processing workflow and integrated with radiometric and physiological measurements, including continuous photosynthetically active radiation (PAR), canopy-level spectral photon flux measurements, photosynthetic photon flux density (PPFD), band-specific photon flux densities, red:far-red ratio (R:FR), stomatal conductance (gₛ), and leaf temperature. PAR measurements beneath the panels were compared with a full-sun control area. Results AVS showed significantly higher shading than AVL (76.14% vs 39.45%, p Conclusion The proposed workflow offers a low-cost, non-destructive tool to quantify shading dynamics and support site-specific assessment of crop microenvironments in agrivoltaic systems. The approach provides crop-relevant information for precision monitoring and management of spatially heterogeneous light conditions across different crop species. Impact The data provided in this manuscript enable the quantification of in-season photovoltaic-induced canopy shading dynamics in an agrivoltaic vineyard using proximal RGB time-lapse imaging and crop-level radiometric measurements. These metrics reflect the spatial and temporal variability of light availability within the vineyard and support site-specific assessment of crop microenvironments and precision management of agrivoltaic systems.
Abstract Context: Accurate leaf area index (LAI) estimation is essential for UAV-based winter wheat growth monitoring, but optical saturation effects under dense canopies remain a major limitation. Aims: This study aims to develop a lightweight dual-stream network (HFI-Net) that improves LAI estimation under high-LAI conditions where saturation effects commonly occur in conventional vegetation indices for efficient UAV-based LAI mapping. Methods: A field dataset containing 637 paired UAV image patches and ground LAI measurements was collected from 21 winter wheat cultivars across seven phenological stages. HFI-Net was proposed, integrating RGB texture and vegetation-index (VI) features derived from multispectral imagery through attention-guided star-shaped multiplicative feature interactions, and five-fold cross-validation with data augmentation was applied. Key Results: HFI-Net achieved a coefficient of determination (\((R^2)\)) of 0.9023 and an RMSE of 0.4822 on the test set. Compared with ResNet50, the proposed model reduced parameters by 98.3% to only 0.40 M, while providing improved prediction performance under saturation-prone high-LAI conditions (\((>4.0)\)). Conclusion: HFI-Net enables reliable winter wheat growth monitoring throughout the growing season with reduced performance degradation under high-LAI conditions and low computational cost. Implications and Impacts: These results indicate the potential of HFI-Net for efficient UAV-based LAI mapping and precision agriculture applications.
Abstract Purpose The study aimed to develop a hybrid Physics-Informed Neural Network (PINN) architecture to forecast the daily strawberry canopy volume using weather data and image-derived inputs. Green pixel counts were used as a low-cost proxy for canopy volume and tracked canopy growth in two strawberry cultivars under real-world field conditions. The PINN architecture was chosen because it directly embeds known growth constraints into the learning process, enabling reliable predictions even with a few observations per day. Methods Two commercially important strawberry ( Fragaria × ananassa Duch.) cultivars, ‘Florida Brilliance’ and ‘FL 16.30–128’ (marketed as Florida Medallion™) were used in this study. Strawberry plant images were acquired at 15-minute intervals over 38 days and were used to train the PINN architecture along with 13 weather data points and accumulated Growing Degree Days (GDD). The PINN was composed of a Long Short-Term Memory (LSTM) residual encoder and a trainable logistic growth curve driven by GDD. Biologically implausible canopy decline during GDD summation increases was penalized by the loss function. Results The Brilliance plant model yielded a Mean Absolute Error (MAE) of 6,996 pixels of canopy size and a Mean Absolute Percentage Error (MAPE) of 4.53%, while the Medallion plant model produced an MAE of 9,440 pixels and an MAPE of 16.4%. Both models outperformed a naive persistence baseline, which produced MAE values of 9,416 and 9,580 pixels for Brilliance and Medallion, respectively, and a physics-only logistic baseline, which yielded an MAE of 10,260 and 10,880 pixels for Brilliance and Medallion, respectively. Conclusion Combining a straightforward growth physics prior with a short-sequence LSTM significantly improved canopy forecasting accuracy, even with about a month of training instances. The system could be used to investigate strawberry growth during the vegetative and early fruiting periods to improve management and increase profit.
Volatile aroma compounds (VACs) are key semiochemicals induced and released by plants under biotic and abiotic stresses. Therefore, it is of great significance for detecting VACs to evaluate the growth environment information and physiological markers of plants. However, real-time and on-site monitoring of VACs at the current stage remains a challenge due to their complex species and trace concentrations. Here, we report an ultrasensitive VAC sensor based on Ru-nanocluster-anchored ZnO nanodisks, for which the corresponding limit of detection for VACs can reach as low as 10 ppb. By constructing temperature-modulated 4 × 4 VAC sensor arrays and integrating artificial intelligence algorithms, we demonstrate simultaneous identification of 10 VAC species and concentrations with 98% accuracy and high fidelity. Furthermore, a portable biomimetic olfactory e-nose based on temperature-modulated sensor arrays was developed to realize the on-site detection of VACs released from tea plants induced by Ectropis obliqua. This work not only provides a design strategy for ultrasensitive VAC sensors but also establishes a scalable platform for real-time plant health monitoring and early warning of biotic and abiotic stresses.
Accurate 3D plant organ segmentation is fundamental to automated phenotyping. Existing approaches rely on annotated training data or species-specific model configurations. We present an annotation-free pipeline for 3D plant organ segmentation, combining text-prompted SAM3 segmentation with semantic neural radiance fields (NeRFs). Given only multi-view RGB images and a list of class names, our zero-shot pipeline produces semantically labeled 3D point clouds without manual annotation, per-species fine-tuning, or domain-specific preprocessing. Multi-view NeRF fusion acts as effective implicit consensus mechanism that lifts imperfect per-frame masks into accurate 3D labels. On a controlled Begonia maculata testbed the SAM3 pipeline achieves 92.6% mIoU, reaching 95.9% of the oracle upper bound established with perfect ground-truth masks. The pipeline was further evaluated on a new dataset spanning ten diverse plant point clouds reaching an average 0.856 mIoU, with leaf and pot IoU above 0.91 and 0.90 for every species, respectively. These results demonstrate that annotation-free 3D plant organ segmentation is now feasible and approaching the range of supervised methods.
Medicinal plants are critical sources of bioactive materials, yet efficient field-scale monitoring of their growth responses to water and nitrogen management remains limited. The present work evaluated Unmanned Aerial Vehicle (UAV) multispectral sensing for estimating leaf area index (LAI), underground fresh weight (UFW), aboveground fresh weight (AFW), underground dry weight (UDW), and aboveground dry weight (ADW) in Astragalus membranaceus var. mongholicus (Bunge) P.K.Hsiao ( A. membranaceus ) under 15 irrigation-nitrogen treatments spanning irrigation inputs of 15.0-30.0 mm per event and nitrogen application rates of 0–450 kg N ha⁻¹. UAV imagery and ground measurements were acquired from 45 plots at 60, 75, 90, 105, 120, and 135 days after sowing (DAS). Thirty-seven vegetation indices (VIs) were screened, and random forest (RF), support vector machine (SVM), back-propagation neural network (BP), genetic algorithm-optimized BP neural network (GA-BP), and decision-level fusion (DLF) models were evaluated with repeated random partitions. Growth responses were trait-dependent. W1N5 reached the peak final LAI (1.96), whereas W2N2 achieved the maximum UFW, AFW, UDW, and ADW (4363.94, 7406.00, 1354.50, and 2596.51 kg ha⁻¹, respectively). VI-trait associations changed with growth stage. GNDVI correlated most closely with LAI at 75 DAS (r = 0.78), and biomass-model performance was stronger during 90–120 DAS than at 60 DAS. DLF was retained in 24 of the 30 stage-specific models. In whole-season modeling, LAI was best estimated by BNDVI input with BP (R 2 = 0.96; RMSE = 0.01), whereas UDW was best estimated by DVI input with DLF (R 2 = 0.93; RMSE = 0.39). Stage-adaptive feature selection increased R 2 by up to 38.60%. Overall, the results indicate that UAV multispectral sensing, stage-adaptive VI screening, and ensemble learning provide a practical framework for monitoring A. membranaceus growth and supporting precision irrigation-nitrogen management in medicinal plant production.
Cadmium contamination severely affects rice growth, yield, and quality, making early stress monitoring essential for agricultural management and food safety. However, traditional detection methods are cumbersome and time-consuming, limiting their applicability to early stress diagnosis. This study developed a rapid and accurate approach for discriminating cadmium stress levels in rice. Arginine-modified flower-like silver nanoparticles (Ag NPs-Arg) were synthesized to enhance Raman signals associated with three stress-response indicators: salicylic acid (SA), malondialdehyde (MDA), and peroxidase (POD) activity. Quantitative prediction models for these physiological indicators and a stress-level discrimination model were established. Among the evaluated models, the CNN-Transformer model achieved the best predictive performance, with Rp 2 values of 0.889, 0.832, and 0.802 for SA, MDA, and POD activity, respectively. An objective weighting method was used to integrate the three biochemical reference indicators, providing a multi-indicator physiological basis for comprehensive stress assessment. The resulting stress-level assessment model achieved an accuracy of 95.83%, demonstrating its ability to capture cadmium-induced physiological changes and assess stress levels in rice.
This study systematically evaluated the contribution of UAV LiDAR structural features such as crop height (CH) and multi-layer gap fraction (GF) and the amplitude of the returning signal represented by normalized intensity (INT), together with multispectral (MS) and thermal infrared (TIR) observations for aboveground biomass (AGB) estimation in winter wheat using a common artificial neural network (ANN) framework. Among the evaluated single sensor approaches, LiDAR features consistently provided the strongest performance, demonstrating the complementary value of crop height, vertically distributed canopy density, and normalized LiDAR intensity for characterizing canopy structure and within-canopy variability. Multi-layer GF improved AGB estimation relative to conventional ground-based GF approaches, highlighting the importance of incorporating the vertical distribution of canopy density. Multi-sensor fusion produced only modest additional improvements, indicating limited benefits relative to the increased acquisition and processing requirements. Temporal analysis showed that structural LiDAR features were most informative during early crop development, whereas normalized intensity, spectral reflectance, and thermal observations became increasingly valuable during canopy maturation and senescence. Comparisons with destructively measured plant area index (PAI), leaf area index (LAI), green leaf area index (GLAI), and green fraction of LAI further demonstrated that normalized LiDAR intensity (903 nm) was more closely associated with green canopy components than purely structural LiDAR metrics. Overall, the results demonstrate that fully exploiting both the structural and spectral information contained within LiDAR observations can substantially improve UAV-based biomass estimation, while multispectral and thermal observations provide complementary information whose contribution varies with crop development and monitoring objectives.
Accurate litchi counting from whole-tree images is essential for yield estimation, orchard management, and plant phenotyping, but remains challenging in real orchards because fruits occur in dense, heavily occluded clusters and vary markedly in scale, illumination, and appearance across ripening stages, particularly when green fruits resemble surrounding foliage. Existing methods have shown promise, but their robustness in complex orchard environments remains limited. To address these challenges, we propose FG-LCNet, a two-stage foreground-guided litchi counting framework. In the first stage, an enhanced fruit-cluster detector improves the localization of small and ambiguous clusters under complex canopy backgrounds. In the second stage, the detected foreground regions are fed into a density-regression network with hybrid attention, while a consistency-based training strategy is introduced to improve robustness to appearance and illumination variations. To support this study, a large-scale litchi counting dataset was established, consisting of 1,126 whole-tree images collected from five orchards and spanning three ripening stages, with approximately 120,000 fruit-level dot annotations and more than 20,000 cluster-level bounding boxes. FG-LCNet achieved the best overall counting performance, with an MAE of 7.44 and an RMSE of 11.01. It showed clear advantages in high-density fruit-cluster scenarios and cross-orchard validation, while maintaining competitive results across orchard-region and maturity-stage subsets. The framework further retained inference efficiency suitable for practical deployment. These results indicate that FG-LCNet provides an effective solution for robust litchi counting and offers potential for other clustered fruit-counting tasks.
Reproduction assets foundThe paper's implementation code is explicitly stated as publicly available at the authors' GitHub repository (FG-LCNet). The litchi counting dataset (1,126 whole-tree images with ~120,000 dot annotations and 20,000+ bounding boxes) is not yet fully public: a ~100-image annotated subset is promised upon acceptance, and,Code · publicdustry Technology Research System (CARS-32-21), Hainan Modern Agricul-
655
tural Industry Technology System (HNARS-08-G02).
656
Conflicts of Interest
657
The authors declare that there is no conflict of interest regarding the publication of this article.
658
Data Availability
659
The implementation code is publicly available at https://github.com/johnhamtom/FG-LCNet .
660
Upon acceptance, a representative subset of approximately 100 annotated litchi images will be
661
released to support reproducibility and preliminary benchmarking. The full dataset is being further
662
organized for future release. Before full release, the complete dataset can be obtained from the
663
corresponding authorOpen asset ↗johnhamtom/FG-LCNetpdf-raw-page:28 lines:1-81Plant phenotyping relevance matchbioRxiv · checked 15 Sept 2026
Water scarcity and increasingly irregular rainfall threaten avocado production in Mediterranean regions, yet the long term physiological responses of mature trees to sustained deficit irrigation remain poorly understood. We conducted a two-year field study integrating continuous monitoring of the soil plant atmosphere continuum, drone-based multispectral imaging, canopy structural analysis, and fruit phenotyping in a mature avocado orchard subjected to three irrigation regimes. The two study years differed markedly in rainfall, providing a unique opportunity to evaluate how environmental conditions modulate tree responses to water limitation. Trees under severe deficit irrigation showed depletion of water in deeper soil layers and a flattened physiological profile, with near-zero diel variation in leaf thickness and trunk water potential, indicating minimal transpiration and decoupling of tree water status from environmental demand. Drone telemetry via NDVI detected stress during fruit growth and maturation, but not during flowering or the new summer leaf flush, revealing greater drought sensitivity at later maturation stages. Although canopy area did not differ among irrigation treatments, canopy surface roughness increased significantly under deficit irrigation, thereby identifying a novel structural indicator of drought stress. Despite large physiological differences among treatments, fruit number remained stable, while fruit weight decreased significantly under severe deficit irrigation, particularly in the wetter year, suggesting that annual rainfall modulates the trade-off between fruit retention and fruit growth. This study provides the first continuous, multi-scale characterization of avocado performance under sustained deficit irrigation in Mediterranean conditions. By integrating plant-based sensors, remote sensing, and artificial intelligence, we reveal previously undescribed stress dynamics and identify new indicators for precision irrigation management in fruit crops.
This study assessed the integration of Unmanned Aerial Vehicles (UAVs) and satellite images (Sentinel-1 and Sentinel-2) in advanced machine-learning techniques to monitor the ABV of quinoa crops (Jacha Grano variety) across the Bolivian Altiplano. The proposed method follows a two-step procedure. First, UAV RGB images were used in photogrammetric and deep-learning (Convolutional Neural Networks-CNN) models to estimate reference quinoa ABVs at a 10 m spatial resolution from the crop canopy 3D model and classification, respectively. Secondly, several spectral and polarization/texture indices derived from Sentinel-2 and -1 images were integrated into three decision-tree-based machine-learning models (Random Forest-RF, Gradient Boosting-GB, eXtreme Gradient Boosting-XGB), and one CNN-based machine-learning model to estimate ABV. Additionally, a Stacking Model (STM) build on top of the three decision-tree-based models was considered for comparison. Model evaluation was also performed in a two-step approach. First, a 10-fold cross-validation strategy was used to highlight ABV sensitivity to Sentinel-2 and Sentinel-1 alone and in combination. Secondly, a Leave-One-Plot-Out Cross-Validation (LOPOCV) strategy was used to avoid autocorrelation between the training and evaluation dataset and therefore provided more insight into ABV mapping potential. The results showed that the combination of Sentinel-1 and Sentinel-2 features in the CNN model achieved the best predictive performance with R2 and RMSE values of 0.64 and 0.39 m3 ∙ 100 m−2, respectively. These findings highlight the potential of integrating multi-source information in advanced artificial intelligence algorithms for quinoa ABV monitoring, offering new insights toward the identification of sustainable practices across remote regions with complex socio-economic contexts.
Plant diseases pose a serious threat to agriculture, causing yield losses of 20 to 40 percent each year, resulting in more than 220,000 million dollars in economic damage and significantly affecting the global food supply. Traditional plant health monitoring practices involve visual inspection of plant tissue and can only detect the presence of disease once visual symptoms are already evident. This article presents AgriIDIA, an early-detection plant disease recognition system trained on datasets of 24-channel multispectral images derived from six optical filters (BlueIR, Hotmirror, K590, K665, K720, and K850) and six vegetation indices (NDVI, GNDVI, NDRE, EVI, REI, and SAVI). First, an exploratory data analysis is conducted on the diagnostic capability of the described 24-channel data representation, using 1,266 image stacks labeled with six classes (diseased/healthy papaya, diseased/healthy potato, diseased/healthy tomato). Next, using the results of the exploratory data analysis, the manuscript describes the training and cross-validation performance of AgriIDIA, with a macro-F1 score of 83.91 ± 3.42% and an accuracy of 84.00 ± 3.17% on the validation set. Finally, the performance of the trained model is evaluated on the reserved test set (N=190), demonstrating an accuracy of 81.05%, a macro-F1 score of 0.7398, and a weighted ROC-AUC of 0.9383. The results of this study suggest that the 24-channel multispectral representation has significant diagnostic potential for the early detection of plant diseases and can be used to design accessible phytosanitary methods for small-scale farmers.
Early detection of tree-seedling establishment is essential for monitoring regeneration success in coastal-dune plantations, where conventional field assessments remain labour-intensive and spatially limited. This study presents a deep-learning workflow for detecting early-stage Pinus pinaster seedlings using multispectral UAS-derived point clouds. Field surveys in the Quiaios National Forest, Portugal, mapped approximately 1500 seedlings using RTK GNSS positioning, biometric measurements, and field photographs. Multispectral imagery acquired with a DJI Mavic 3 Multispectral platform was processed through Structure-from-Motion to generate calibrated orthomosaics, terrain products, and dense point clouds. Training-data preparation combined pine-centred buffers, spectral conditioning, manual refinement and point-cloud class assignment. Point Transformer V3 models were trained in ArcGIS Pro and evaluated using field-mapped buffers withheld from model training within plantation-line areas. The Baseline high-recall model achieved 88% object-level recall at the operational threshold of at least three classified Pine-Seedling points per buffer. The refined hard-negative model retained 84% recall while reducing off-buffer detections from 243 to 41. False-negative analysis showed that omissions were associated with reduced crown diameter and limited branch development under the adopted buffer-based retrieval framework. These results support transformer-based multispectral point-cloud classification for scalable monitoring of early-stage pine regeneration in heterogeneous coastal environments.
Occlusion is a major factor limiting accurate three-dimensional (3D) wheat phenotyping. In natural growth conditions, overlapping spikes, leaves, and stems often make only partial target regions visible in single-view images, hindering complete and reliable 3D reconstruction. To address this problem, this study proposes an amodal completion-assisted, sequential framework for single-view 3D reconstruction of occluded wheat. The framework first uses visible prompts to recover the complete appearance and structural cues of occluded targets, and then feeds the completed images into single-view 3D reconstruction models to generate complete 3D structures. We construct the MMWO (Multi-view Multi-instance Wheat Occlusion) dataset from MMW, which is captured under controlled indoor scenarios, by synthesizing diverse occlusion samples through organ-level cutouts, random geometric transformations, and region-constrained pasting, with annotations including visible masks, occlusion masks, and complete target images. Six representative reconstruction methods, including Direct3D, Real3D, SF3D, Spar3D, TRELLIS.2, and Hunyuan3D, are systematically evaluated. Hunyuan3D achieves the best geometric performance, with the lowest mean CD- \(L_1\) and CD- \(L_2\) values of 0.1286 and 0.0536, and the highest mean F-score of 0.5668. SF3D achieves the best rendering quality in terms of PSNR, SSIM, and LPIPS. In addition, Pix2Gestalt completion reduces the estimation errors of spike length, width, and area from 9.31%, 10.89%, and 32.23% to 4.64%, 9.70%, and 9.45%, respectively. These results demonstrate that amodal completion can effectively alleviate occlusion-induced information loss and provide more complete structural priors for single-view 3D wheat reconstruction. This study offers a feasible solution for 3D wheat phenotyping under occlusion and provides a systematic reference for applying 3D generative models to agricultural phenotyping.
Miho H, Barranco Navero D, Trujillo Navas I, Morello Parra P, Koubouris G, Trapero Ramirez C, López-Bernal Á, Oueslati A, Flores Mesas D, García López MT, Valverde Caballero P, Deger RE, Meca E, Perri E, Santilli E, Moral Moral J, Estudillo Cazorla C, Cabello Pozo D, Yousef Yousef M, Acar S, Gurbuz-Veral M, Priego Capote F, Zahiri A, Oulbi S, Márquez Pérez MI, M. Díez C.
Background Olive ( Olea europaea L.) breeding initiatives rely heavily on the extensive and correct characterisation of genetic resources to successfully achieve their goals, such as addressing climate change and emerging diseases challenges. However, the historical lack of standardised phenotyping protocols across multi-environment trials has severely hindered data interoperability and large-scale comparative analyses. Methods Within the Horizon 2020 GEN4OLIVE project, five international olive germplasm banks established a consensus-based methodological framework to systematically evaluate over 500 olive cultivars. We harmonised 14 evaluation protocols covering six fundamental dimensions: phenological and agronomic traits, abiotic stress resilience, biotic stress resilience, olive oil yield and chemical quality, table olive quality assessment, and morphological characterisation and photography. While most protocols were adapted from previously published literature to ensure ease of implementation across different facilities, novel methodologies for frost tolerance and standardised photography were developed de novo. Results The implementation of these consensus methods across five countries proved highly successful. This methodological framework enabled the generation of the largest harmonised, publicly available dataset on olive genetic resources to date, effectively making possible the correct comparation and ranking of the olive cultivars based on their specific characteristics. Conclusions This compendium of methods provides a robust, highly replicable reference point for the standardisation of olive germplasm characterisation and use of shared benchmark cultivars as an effective way for data normalization and comparation. It facilitates future global pre-breeding efforts, ensures international data interoperability, and supports the discovery of resilient cultivars to secure the future of the olive sector.
Reproduction assets foundThe article declares two paper-specific public assets: the GEN4OLIVE phenotypic dataset from evaluating over 500 olive accessions across five germplasm banks, hosted on the project's Olive Varieties Database, and a Zenodo-deposited methodological handbook (Extended Data) containing the 14 protocols, visual assessment, Dataset · publicData and software availability
The phenotypic dataset generated from the evaluation of over 500 olive varieties across the five Mediterranean
germplasm banks using this compendium of protocols and methodologies, is publicly available via the GEN4OLIVE
project repository.
• Repository: GEN4OLIVE Olive Varieties Database.
• Link: https://www.uco.es/ucolivo/gen4olive/olivevarieties (GEN4OLIVE Database, 2025).
Page 8 of 15
Open Research Europe 2026, 6:322 Last updated: 14 SEP 2026Open asset ↗GEN4OLIVE Olive Varieties Databasepdf-raw-page:8 lines:1-44Plant phenotyping relevance matchOpenAlex · checked 15 Sept 2026
Modern crop breeding demands precise organ-level analysis for trait quantification, making plant point cloud segmentation (PPCS) increasingly important. However, conventional deep learning approaches rely heavily on densely annotated datasets that are labor-intensive to acquire. Unified PPCS adaptation from distribution-shifted examples with minimal additional training remains challenging. To address this, we propose PlantC2USeg, a deep transfer learning framework featuring cross-scale consistency learning to explicitly align features across spatial scales and an information-restricted decoding strategy that prevents reconstruction shortcuts and promotes robust adaptation. The resulting pre-training enables stable few-shot generalization across species and sensing conditions, while unified fine-tuning with inherited thresholds further reduces adaptation overhead. Under full supervision on Soybean3D, PlantC2USeg achieves the highest semantic IoU and instance mWCov among compared methods, at 91.91% and 94.62%. With 20 labeled samples, it leads both metrics at 89.78% and 90.27%; with only 10 samples, it retains the highest mWCov of 83.23% while achieving 83.19% IoU. Across HR3D, 10-shot transfer to tobacco, tomato, and sorghum averages 78.41% IoU and 79.42% mWCov, while 22-shot transfer to SYAU-Maize achieves the highest IoU and mRec at 92.75% and 93.51%. Furthermore, a leading category-averaged mIoU of 85.0% on ShapeNet Part demonstrates the framework's capability to handle diverse shape variations beyond agricultural domains. These results demonstrate that PlantC2USeg reduces overall adaptation effort under distribution shifts, enabling scalable plant phenotyping and transferable 3D representation learning beyond agriculture.
To enhance crop performance, intercropping strategies leverage volatile organic compound (VOC)-driven interactions with companion plants that constitutively emit VOCs. Despite the agricultural importance, the mechanisms and kinetics of VOC-mediated sensory transduction in receiver plants eavesdropping on neighbouring non-kin emitters remain largely unknown due to a lack of appropriate non-destructive analytical tools. In this work, we employ multiplexed salicylic acid (SA) and H 2 O 2 nanosensors in Brassica rapa subsp. Chinensis (pak choy) plants to visualize, in real time, reactive oxygen species (ROS) and SA signal transduction following exposure to constitutively-released VOCs from neighbouring aromatic plants-namely sweet basil and spearmint. Unique emitter-specific temporal signatures of ROS and SA were observed in receiver pak choy: sweet basil VOCs induced concomitant generation of ROS and SA at 30 min, whereas spearmint VOCs triggered SA production at 30 min, followed by ROS accumulation. The temporal data enabled the formulation of a diffusion model that quantifies the VOC perception threshold that triggers the distinct early ROS and SA signalling. Transcriptomics analysis at 2 h revealed that both emitters evoke largely distinct changes in pak choy, likely stemming from variations in speed and sequence of the early signal transduction, leading to different phenotypic outcomes. Intercropping with sweet basil led to enhanced pak choy biomass, stress resilience and secondary metabolite accumulation, whereas spearmint as companions had a limited impact. Our study captures in real time, the VOC-induced rapid signalling in receiver plants and its ensuing effect on growth. These nanosensor-enabled findings represent an important advance in deciphering how emitter-specific volatile cues are integrated into plant responses, guiding rational selection of beneficial companion plants for improved yield and nutritional profiles in sustainable agriculture.
Agriculture is being revolutionized through the integration of cutting-edge technologies that support efficiency, productivity, and sustainability. Recent trends in transforming traditional agriculture into smart agricultural systems through the application of Artificial Intelligence (AI), IoT, and drone technology are gaining popularity worldwide. In this context, the current study developed an integrated artificial intelligence-based farming system incorporating Internet of Things sensors, drones, deep learning, and web/mobile applications for timely monitoring of crop and aquatic health. Data were collected using IoT sensors and drones equipped with high-resolution cameras to monitor soil and water parameters, including temperature, humidity, and pH, under various climatic conditions. A dataset of 70,000 images was used for system training, validation, and testing with an 80:10:10 split, together with three months of IoT sensor data. Models including YOLOv8, CNN, Faster R-CNN, ResNet50, MobileNet, EfficientNet, DenseNet, LSTM, and Random Forest were used for pest and disease detection, fish classification, fish disease detection, shrimp disease detection, and monitoring of climatic factors. Data pre-processing included denoising, normalization, missing-value handling, and feature extraction. The designed model showed reliable performance across the tasks, with 88.4% accuracy for shrimp detection, 92.1% for pest detection using YOLOv8, and 93.4% accuracy for plant disease detection using the CNN model. Overall performance was recorded at 97% accuracy, with high precision, F1-score, and mAP, and an RMSE of 1.2 for sensor-based prediction and validation. The current findings indicate the potential of using AI, IoT, and drone technologies to detect biotic and abiotic stresses during farming and support a sustainable agricultural system.
Key message Yield-Graph enables accurate maize yield prediction from incomplete multi-stage phenotypic and environmental data by modeling higher-order environment-trait interactions, with robust applicability across growth stages, regions, and crop species. Accurate yield prediction before maize harvest is crucial for advancing agricultural management and ensuring food security. Unlike conventional approaches that rely on traits from a single growth stage, this study models multiple traits across different developmental stages, all targeting final yield, thereby uncovering their stage-specific contributions and demonstrating the feasibility of early yield prediction. We introduce Yield-Graph, an innovative framework that evaluates phenotypic data at distinct developmental stages for yield prediction. The method employs a bipartite graph structure to impute missing trait values at each stage and leverages a hypergraph attention mechanism to capture high-order sample relationships. Comprehensive benchmark experiments demonstrate that Yield-Graph matches the top-tier predictive accuracy of exhaustively optimized tree models. Moreover, the framework exhibits strong robustness across growth stages, high adaptability to regional variations, and effective generalization across datasets. These findings highlight the potential of graph-enhanced multi-stage modeling for early-stage yield prediction, offering a scalable solution for precision agriculture and intelligent crop management.
High-throughput phenotyping of lettuce seedlings is highly prone to background confusion because the seedlings are small, have weak textural features, and exhibit spectral reflectance similar to that of the substrate. Traditional single-visual-modality approaches struggle to achieve reliable structural and physiological characterization simultaneously under the repetitive backgrounds and dense arrangements typical of greenhouse tray cultivation. To address these challenges, we establish a multimodal 3D phenotyping framework tailored for controlled agriculture environments, enabling the quantification of structural and physiological characteristics of lettuce seedlings. This framework is based on an unmanned ground vehicle (UGV) platform integrating a RGBD camera and a quad-band multispectral sensor which are rigidly coupled and synchronously triggered. An alignment module based on established feature matching algorithm is introduced to register the misalignment between source multispectral and RGBD images. Subsequently, we design a novel dual-backbone instance segmentation network, MS-SegNet, to enhance segmentation accuracy by hierarchically fusing geometric information with multispectral features. A robust 3D metric pose estimation pipeline, incorporating standard SfM initialization, scale recovery, and generalized ICP refinement, is constructed to generate 3D point clouds with spectral attributes and semantic labels. Finally, key structural and physiological phenotype parameters of each seedling are calculated based on the 3D semantic multispectral point clouds. Experiments demonstrate that MS-SegNet achieves significant advantages in instance segmentation of lettuce seedlings with mAP@50:95 = 0.854. The metric 3D pose estimation pipeline exhibits reliable performance under complex controlled conditions. The quality of the 3D reconstructions is indirectly validated through downstream structural trait extraction. The estimated seedling height and crown width show high correlation with manual measurements, achieving R 2 values of 0.8379 and 0.918, and RMSE values of 10.94 mm and 11.56 mm, respectively. Overall, by systematically integrating these adapted components with the novel segmentation architecture, this framework achieves stable performance improvements in 3D reconstruction, instance segmentation, and phenotypic analysis under greenhouse conditions. It provides a scalable, integrated technical solution for non-destructive, high-throughput phenotyping of crop seedlings in controlled environments.
Accurate plant disease detection remains challenging when using single-modality data, which fails to capture comprehensive disease-related features. However, many existing studies rely on pixel-level classification or prior plant segmentation and lack explicit modeling of cross-modal interactions, limiting their ability to distinguish between healthy and diseased plants. This study presents a spatial–spectral fusion deep learning framework (S2-PDD) that fuses uncrewed aerial vehicles (UAV)-based RGB imagery and hyperspectral-derived feature representations for plant-level detection of potato diseases (blackleg and Potato virus Y (PVY)), employing early fusion (modalities combined at the input stage) and middle fusion (features integrated at intermediate stages within the model backbone) strategies. The multimodal fusion models were compared against single-modal models and an existing S2ADet model. Model performance, assessed through five-fold cross-validation, demonstrated that multimodal models integrating RGB and vegetation index features achieved the highest mAPs of 86.65 ± 1.71 (%; E-RV model) and 85.74 ± 1.96 (M-RV), respectively. These mAPs were higher than those of all single-modal models, including the RGB-only (83.21 ± 1.46) and hyperspectral-only models (PCA features: 79.71 ± 1.45; vegetation index features: 85.31 ± 2.36). They also exceeded mAPs of multimodal models combining RGB with PCA features (early fusion: 83.00 ± 2.81; middle fusion: 83.11 ± 2.46; S2ADet: 84.04 ± 2.73), regardless of the fusion strategy. The superior performance highlights that vegetation index features provide strong class separability compared to other hyperspectral representations. The proposed models achieved strong plant-level detection performance, with AP of 78.65 ± 4.14 (E-RV) and 77.40 ± 3.46 (M-RV) for blackleg disease, as well as 84.82 ± 3.59 (E-RV) and 83.28 ± 3.86 (M-RV) for PVY. These results demonstrate the potential of UAV-based multimodal sensing for disease monitoring in cropping systems. A potato plant disease detection dataset was constructed and made publicly available, containing paired RGB and hyperspectral image tiles with bounding box annotations. The code is available at https://github.com/Tim-Agro/S2-PDD.
Accurate and efficient monitoring of tea plant growth parameters via remote sensing is essential for precision plantation management. However, spectral indices relying solely on reflectance often exhibit limited sensitivity in capturing complex tea canopy characteristics. This study developed a data-driven framework integrating spectral reflectance, frequency-domain harmonic components, and spatial texture features to construct tri-feature fusion indices (TFIs) and establish machine learning and deep learning models for tea growth monitoring. Ten-band multispectral imagery was acquired using a UAV alongside synchronous field measurements of leaf and plant biomass and nitrogen accumulation. TFIs were constructed through exhaustive feature combinations and optimized via a data-driven search strategy. Subsequently, random forest (RF), multilayer perceptron (MLP), convolutional neural network (CNN), and transformer models were evaluated using a leave-one-site-out cross-validation (LOSO-CV) strategy. The selected TFIs showed strong associations with tea growth parameters within the investigated dataset, with R2 values up to 0.63 and 0.62 for leaf dry matter and leaf nitrogen accumulation, respectively. Models incorporating selected TFIs achieved cross-validated R2 values of 0.56 for leaf dry matter (MLP), 0.59 for plant dry matter (MLP), 0.73 for leaf nitrogen accumulation (MLP), and 0.68 for plant nitrogen accumulation (CNN). These models exhibited competitive predictive performance comparable to RF, although no statistically significant differences in mean absolute error were observed under site-held-out evaluation. Furthermore, model-derived spatial maps provided insights into fine-scale spatial heterogeneity and potential interannual variations in tea growth parameters across representative plantations from 2024 to 2025. Overall, this study provides a UAV-based framework for tea growth parameter estimation by integrating multi-domain information without requiring additional environmental observations.
Fruit quality is a critical determinant of economic returns in pear production, and maintaining an appropriate fruit load (FL) is essential for achieving high yield and quality. As a direct indicator of canopy photosynthetic capacity and assimilate supply, leaf number constitutes the key biological basis of reasonable FL determination under the leaf-to-fruit ratio concept. However, accurate and efficient estimation of leaf number in mature pear trees remains technically challenging, limiting its practical use in precision FL regulation. Here, we propose a data-driven framework for leaf number and reasonable FL estimation by integrating 3D point cloud-derived canopy structure with machine learning. A pipeline for extracting 3D architectural traits was developed and implemented in the software tool FTPCT, enabling rapid and standardized trait acquisition. Through correlation analysis, multicollinearity diagnosis, and variance inflation factor screening, five key traits strongly associated with leaf number were identified and incorporated into five machine learning models optimized using Bayesian optimization. Among them, the optimized random forest regression model achieved the highest and most stable performance, with R 2 of 0.85, RMSE of 239.74, and MAE of 149.26 for test dataset. SHAP analysis identified tree crown volume as the dominant contributor to leaf number estimation. Field validation demonstrated that FL regulation guided by the proposed framework significantly improved fruit weight and size without reducing yield compared with conventional practices. Notably, the proposed approach avoids explicit leaf-level reconstruction and relies on less canopy-scale traits, substantially reducing data requirements and computational cost, and thereby offering strong potential for rapid, field-deployable FL regulation in large-scale orchards.
Reproduction assets foundThe paper's phenotyping analysis assets are the authors' publicly released LeafNumPred source code and trained models, and the FTPCT software for 3D trait extraction from pear tree point clouds. Phenotype/point-cloud datasets are only available on request.Code · public. Supplementary data
The following is the Supplementary data to this article:
Multimedia component 1
mmc1.docx (1.6MB, docx)
Data availability
Data will be made available on request. Anyone who wants to obtain other public data can contact us at taost@njau.edu.cn. The source codes and models have been made publicly available at https://github.com/Zhang-Fanhang/LeafNumPred, and the FTPCT software has been released at https://github.com/Zhang-Fanhang/FTPCT/tree/Installation-package.
References
1.Tao S., Khanizadeh S., Zhang H., Zhang S. Anatomy, ultrastructure and lignin distribution of stone cells in two Pyrus species. Plant Sci. 2009;176:413–419. [Google Scholar]
2.Zhang F., Wang Q., Yuan K.Open asset ↗Zhang-Fanhang/LeafNumPredhtml-lines:284-315Code · publical variations [34,35]. The method for calculating these traits are shown in the Supplementary information 1.
2.5. Software implementation for 3D trait extraction (FTPCT)
To facilitate efficient and standardized extraction of canopy structural traits from point cloud data, we used a standalone software tool, FTPCT (available at: https://github.com/Zhang-Fanhang/FTPCT/tree/Installation-package), which integrates the trait extraction procedures applied in this study. The software provides a graphical user interface, enabling users to process tree-level point cloud data and extract key 3D structural traits without requiring advanced programming skills.
FTPCT implements a series of predefined proOpen asset ↗Zhang-Fanhang/FTPCThtml-lines:138-149Plant phenotyping relevance matchCrossref · checked 14 Sept 2026
Problem: Agriculture plays a pivotal role in the Indian economy, where crop production quality and quantity directly impact the livelihoods of millions. Climate variability, emerging plant diseases, and improper pesticide application contribute significantly to agricultural losses. Early and accurate disease detection is crucial for mitigating crop damage and ensuring food security. Methodology: This study presents a novel deep learning framework based on EfficientNet architecture, enhanced with Progressive Fine-Tuning Strategy for automated plant disease detection. The proposed methodology was evaluated on two benchmark datasets: the Plant Village dataset comprising 20,639 images of tomato, potato, and bell pepper with 15 disease varieties; and a drone-captured rice plant dataset containing 4432 samples from public repositories. The model’s performance was assessed using multiple metrics, including classification accuracy, precision, recall, and F1-score. To strengthen the validation of high accuracy results, additional statistical analyses like class imbalance ratio, Entropy, Chi-Square test, convergence curve, ANOVA test, Tukey’s post hoc HSD test, confidence interval, Cohen’s Kappa result, fold-wise dispersion analysis, mean, std deviation are included in the manuscript. Result: Experimental results demonstrate that the fine-tuned EfficientNetV2-B1 architecture achieved exceptional performance with 99.7% classification accuracy on the PlantVillage dataset and 99.03% accuracy on the drone-based rice disease dataset, significantly outperforming existing state-of-the-art transfer learning models. Model explainability techniques further validated the reliability and interpretability of the predictions, confirming the model’s focus on disease-relevant features.
Aboveground biomass density (AGBD) is a key indicator in agricultural systems, directly reflecting crop carbon storage potential and yield levels. The World Food Studies (WOFOST) model is widely used for crop growth simulation due to its process-based interpretability. However, its application is limited by complex calibration demands and struggles with spatial heterogeneity. To address these limitations, this paper proposes an assimilation system that integrates Synthetic Aperture Radar (SAR) data into a modified WOFOST model, which requires simple calibration. WOFOST is run in potential production mode, which assumes ideal conditions to reduce input data requirements. Before assimilation, phenology-related temperature sums are aligned with phenology and meteorological data to match local growth stages. Two quantities are then estimated and updated in the model through data assimilation by minimizing the difference between SAR-derived and simulated AGBD. These quantities are the proposed yield reduction factor, which represents the proportional decrease in potential CO 2 assimilation under stresses, and the initial total dry weight at sowing. Validation experiments were conducted using multi-year cotton datasets from two farms in Georgia, USA, differing in whether irrigation was applied. Compared to WOFOST simulations and evaluated against in situ AGBD measurements, the assimilation results improve agreement and reduce error (approximately 43% RMSE reduction at the rainfed site and 15% at the irrigated site). It also delivers spatial maps of biomass, together with model-derived yield and harvest-index diagnostics, and remains operational under frequent cloud cover where optical observations are sparse. This SAR-based assimilation strategy reduces calibration demands, providing a novel and practical pathway to extend WOFOST applications to diverse agricultural scenarios.
Wu F, Liu S, Zohner CM, Townsend PA, Crowther TW, Peñuelas J, Yang D, Yang N, Dong T, Xu W, Wang Z, Liu X, Dai G, Dong J, Durán SM, Schneider FD, Zeng Y, Cornelissen JHC, Kattge J, Wu J, Asner GP, Cavender-Bares J, Reich PB, Yan Z.
Global trait axes reveal overarching dimensions of plant functional variation. However, how these dimensions are spatially organized within and across forest types remains unclear. We combined drone-based full-range imaging spectroscopy with crown-level measurements of 16 physiological, morphological and biochemical traits across temperate, subtropical and tropical forests in China to enable spatially-explicit trait mapping. Through site-training scenario, leaf-to-canopy scaling and spectral-domain modelling tests, we find that reliable canopy trait retrieval depends not only on trait and spectral coverage, but also on preserving trait-spectral relationships across sites and scales. Spectral predictions recovered observed multivariate covariation, summarizing crown variation into a leaf-economics dimension and two additional biochemical dimensions related to hydro-thermal regulation and defence/metabolism. Mapping these dimensions revealed distinct community-level trait organization alongside substantial species- and crown-level variation within forests. These findings link remotely sensed trait retrieval to environmental filtering and plant functional differentiation, providing a scalable framework for monitoring forest functional diversity.
Grain yield is a central target in wheat breeding, yet accurately predicting it remains challenging because it depends on many genes and responds strongly to environmental variation. Genomic selection (GS) has improved breeding efficiency by enabling genome-based prediction of genetic merit, but predictability (PA) for grain yield is often limited under stress environments. At the same time, advances in high-throughput phenotyping (HTP) using unmanned aerial vehicles (UAVs) provide phenomic data that capture environment-responsive plant performance and may complement genomic information. In this study, we evaluated genomic and phenomic models for predicting grain yield in elite bread wheat lines across irrigated, drought, and heat-stress environments. Using a sparse phenotyping framework, we compared parametric and non-parametric models. PA was evaluated within environments and under cross-environment sparse phenotyping scenarios. Genomic models provided a stable baseline and enabled effective information sharing across environments when phenotypic data were incomplete. Phenomics-only models captured environment-specific plant responses but were more sensitive to environmental context. Multiomics models that integrated genomic and phenomic information consistently achieved the highest PA, with the largest gains observed under stress conditions. Overall, our results demonstrate that integrating genomics and UAV-based phenomics within sparse phenotyping designs offers a practical and scalable approach to improve grain yield prediction in wheat.
Reproduction assets foundThe paper's grain yield BLUEs, spectral wavelength BLUEs, and genotypic data are publicly deposited in the CIMMYT data repository (https://doi.org/10.71682/10549399), directly reproducing this paper's phenotyping measurements. No author analysis code with a public URL is stated; other URLs are generic tools/services.Dataset · publicok.com.
Paolo Vitale, Email: p.vitale@cgiar.org.
DATA AVAILABILITY STATEMENT
The datasets generated and analyzed during this study, including best linear unbiased estimates (BLUEs) for grain yield and spectral wavelengths, as well as the corresponding genotypic information, are publicly available in the CIMMYT data repository ( https://doi.org/10.71682/10549399 ).
REFERENCES
Araus, J. L.
,
Kefauver, S. C.
,
Zaman‐Allah, M.
,
Olsen, M. S.
, &
Cairns, J. E.
(2018). Translating high‐throughput phenotyping into genetic gain. Trends in Plant Science, 23(5), 451–466. 10.1016/j.tplants.2018.02.001
Brault, C.
,
Lazerges, J.
,
Doligez, A.
,
Thomas, M.
,
Ecarnot, M.
,
Roumet, P.
,
Bertrand, Y.Open asset ↗CIMMYT data repository · 10.71682/10549399lines:280-433Plant phenotyping relevance matchEurope PMC · checked 14 Sept 2026
Whitefly-transmitted begomoviruses cause tomato leaf curl disease (ToLCD). In India, at least 15 begomoviruses are known to cause ToLCD, posing a major challenge to resistance breeding. Although several Ty resistance loci have been introgressed from wild tomato relatives, variable resistance responses are frequently observed, likely due to mixed infections and the absence of a standardized disease scoring system. Moreover, limited knowledge of the infecting begomoviruses in resistant genotypes has hindered the effective use of donor lines in breeding programs. This study evaluated 17 homozygous Ty-gene donor tomato genotypes under natural epiphytotic conditions and identified the associated begomoviruses. To quantify disease severity, a robust disease scoring system was developed using coefficient of infection (CI) by integrating symptom parameters- leaf curling, leaf smalling, stunting, and fruiting, with population-level disease incidence. Field evaluations for two years revealed that genotypes carrying both Ty-2 and Ty-3 loci showed higher resistance, though variability existed among them. Genotypes with Ty-3 alone or Ty-5 + Ty-6 combinations also displayed substantial tolerance, and five genotypes were identified as highly resistant. Molecular indexing revealed frequent mixed infections and identified multiple begomoviruses, including a newly characterized species, tomato leaf curl Ty Pusa virus, alongside tomato leaf curl New Delhi virus, tomato leaf curl Palampur virus, tomato leaf curl Gujarat virus, and tomato leaf curl Joydebpur virus. These findings highlight a shift in begomovirus predominance and possible recombination-driven emergence of new variants. This study provides an integrated framework for evaluating ToLCD resistance and emphasizes the need for continuous reassessment of resistance sources to ensure durable tomato cultivar development.
Alexander Bucksch · Yong S. Chung · Jennifer Clarke · Stephan Gerth · Philipp von Gillhaußen · Wei Guo · Jana Kholová · Shree Pariyar · Ethan Pickering · Sindhuja Sankaran · Sahameh Shafiee · Sergio Alan Cervantes-Pérez · Stijn Dhondt · Zhiguo Han · Kabir Hossain · William LaVoy · Jonathan P. Lynch · Sonia Negrão · Tony Pridmore (427291) · Hannah Schneider · Stefan Schwartz · Ian Stavness · Shangpeng Sun · Vadez · Lee West · H.J. van de Zedde
Field / plotWhole plant / canopy / plot / field
Plant Phenomics studies the phenotypic dynamics that form plant phenotypes by systematically measuring traits using phenotyping methods from the quantum to the ecosystem levels. It has emerged as an interdisciplinary field advancing sensing, computation, and plant biology. However, Plant Phenomics has lacked a unifying framework that integrates its community's core concepts from the formal and life sciences. Central to this framework is the definition of the phenome as a set of phenes that govern phenotypic dynamics across all spatial and temporal scales of biological and ecological organization and in interaction with the environment. This paradigm moves beyond gene-centric views and recognizes the equal importance of all spatial and temporal scales in forming plant phenotypes, advancing Plant Phenomics as a data-driven discipline and its emerging profession, the plant phenomicist.
Year-to-year climate variability poses a challenge for agriculture by increasing crop yield variability; therefore, there is a need to identify genotypes that can withstand these fluctuations. With the right selection criteria, genotypes with yield stability across variable environmental conditions can be selected. Methods such as Finlay-Wilkinson random regression (FWRR) may allow us to use sparse datasets-common in plant breeding pipelines-and incorporate genomic data to leverage phenotypic information from related genotypes to predict yield stability. Our objective was to examine how the number of environments and the variance among those environments affect stability predictions. We also integrate FWRR as a genomic prediction tool for characterizing yield stability, comparing it to the traditional genomic prediction models as a reference. We used three datasets: one highly unbalanced dataset for oats (Avena sativa L.) and two completely balanced datasets with different numbers of environments for barley (Hordeum vulgare L.) and wheat (Triticum aestivum L.). We fit standard Finlay-Wilkinson (FW) and FWRR models to estimate grain yield and stability under various scenarios. We found that the estimated stability values obtained were similar using balanced datasets for FW or FWRR. FWRR also achieved moderate predictive ability for stability using unbalanced datasets under 10-fold cross-validation (CV1) with new genotypes. In terms of environmental representation, selecting the right set of environments for inclusion in the model was more important than adding more environments. Our results suggest the possibility of using FWRR to select stable genotypes earlier in line development, as well as to design resource-efficient stability-testing schemes.
Salt stress can markedly alter seedling architecture, creating a need for non-destructive three-dimensional (3D) phenotyping methods capable of resolving fine plant structures. However, organ-level segmentation of plant point clouds remains challenging because of leaf overlap, slender stems, ambiguous stem–leaf boundaries, and severe class imbalance. In this study, we developed PTV-SegCo, a task-adapted Point Transformer model for organ segmentation and structural phenotyping of coriander seedlings under salt stress. PTV-SegCo integrates efficient channel attention, gated shallow–deep feature fusion, and a combined cross-entropy–Dice loss to improve representation of fine and minority organ structures. The dataset comprised 60 manually annotated 3D point-cloud samples from 12 cultivation trays repeatedly observed over five acquisition dates under six NaCl concentrations (0, 50, 100, 150, 200, and 250 mmol L −1 ). Because the earliest acquisition represented a particularly challenging developmental stage, these 12 samples were used as a fixed early-stage model-selection set, while samples from the remaining four dates were organized into four date-blocked training–validation configurations. Under this internal model-development protocol, PTV-SegCo achieved mean mAcc and mIoU values of 93.05% and 89.05%, respectively, and showed numerically higher performance than its direct backbone PTV-Seg50. These values should be interpreted as internal comparative results rather than as an unbiased estimate of generalization to unseen cultivation trays, and the present results should not be interpreted as establishing the broad competitiveness of PTV-SegCo against other point-based, convolution-based, graph-based, transformer-based, or plant-specific segmentation architectures. After semantic segmentation, reconstructed scenes were metrically calibrated using the known cultivation-tray dimensions, followed by individual-plant separation and quality control. Four reconstruction-derived structural descriptors—plant height, projected area, voxel occupancy volume, and leaf point ratio—were extracted to characterize temporal structural variation under different NaCl treatments. For treatment-level inference, individual-plant measurements were aggregated within each cultivation tray at each acquisition time, with the cultivation tray treated as the independent experimental unit. Independent manual validation showed close agreement for plant height and projected area, with R 2 values of 0.9969 and 0.986, respectively. Overall, the proposed workflow provides a feasible approach for organ-level segmentation and automated 3D structural analysis of small coriander seedlings under salt stress. The extracted descriptors primarily represent reconstruction-derived spatial characteristics and should not be interpreted as direct indicators of physiological status; voxel occupancy volume and leaf point ratio remain without direct external validation.
ABSTRACT Water-use efficiency (WUE), the ratio of accumulated plant biomass to water lost through transpiration has conventionally been determined using a destructive single-point measurement. Recent advances in high-throughput phenotyping now enable repeated, non-destructive estimation of biomass and WUE. However, these digital measurements must be statistically validated against conventional destructive methods to validate their use as reliable proxies. Therefore, we compared digital biomass determined point clouds produced from multispectral camera scanners with destructive harvests across eight harvests using Samsun tobacco grown under both drought and high-water conditions. WUE efficiency, calculated using the digital biomass estimated from a point cloud and gravimetric water use determinations, were compared to destructive harvest determinations. The coefficient of variation (CV) showed there were no significant differences in digital and destructive measurements for either biomass or WUE. Indicating that digital measurements can be used in place of destructive measurements. Drought plants used significantly less water and were significantly smaller than high-water plants from Harvests 4 through 8. However, there were no significant differences in the ratio of evapotranspiration to leaf area or WUE, indicating that drought plants were simply smaller and used less water than the high-water plants. This work validates that estimating plant biomass from a digital point coupled with continuous gravimetric determination of water use provides a reliable nondestructive measure of WUE in high-throughput measurements across the full plant life cycle. PLAIN LANGUAGE SUMMARY We grew tobacco plants under either a drought or high-water treatment and harvested a portion of the plants every few days for a total of eight harvests. Throughout the experiment, we collected 3D images of the plants and continuously measured pot weight to track plant growth and water use across different developmental stages. Destructive biomass served as the gold-standard measurement. We then compared biomass and water-use estimates generated from the digital measurements with the destructive measurements. The digital approach provided accurate estimates of plant biomass and water use while requiring little hands-on labor and no plant destruction. These nondestructive methods could help plant breeders identify water-efficient plants earlier in the breeding process, accelerating the development of crops that use water more efficiently.
Abstract Tree crowns are complex, three‐dimensional structures whose morphology varies among species, individuals and environments. Although light detection and ranging (LiDAR) provides high‐resolution, single‐tree point clouds that advance species discrimination and the assessment of intraspecific variation in situ, crown shape is still commonly reduced to low‐dimensional metrics (e.g. crown diameter or crown base height), losing much of its three‐dimensional geometric complexity. We introduce a fully 3D geometric morphometric framework that captures crown shape directly from LiDAR point clouds at both species and individual levels. Pre‐segmented LiDAR single‐tree point clouds of eight temperate forest species were converted into three‐dimensional shape representations using radial bounding volumes (RBVs), which partitioned each crown into a standardized set of vertical layers and radial sectors. Surface points automatically digitized from each RBV formed geospatially aligned, 3D pseudolandmark configurations representing geometric morphometric crown shapes. These configurations served as the input data for multivariate analyses of crown shape variation within and between species. Twelve structural traits, including crown and stem dimensions, were extracted from the same RBVs and integrated into analyses of trait–shape associations. The morphospace of crown shape was structured along different axes of variation in broadleaf species than in conifers. Within these groups, species pairs—such as Fagus versus Quercus and Picea versus Pinus —exhibited contrasting intraspecific morphological gradients, with different structural traits driving shape variation in each. Crown base height and total crown height emerged as the strongest predictors of crown shape. Differences in crown shape among species were primarily captured by symmetric components, with asymmetry providing a negligible signal. Interspecific differentiation was largely driven by architectural variation rather than pure size differences. Morphological differences derived from pseudolandmarks and convolutional neural network features exhibited stronger correlations in conifers than in broadleaf species. We present a reproducible, LiDAR‐native framework for quantifying and comparing 3D crown morphology within and across species. Using the RBV approach, geospatially aligned pseudolandmarks can be derived from any pre‐segmented, single‐tree LiDAR point cloud, enabling scalable, multi‐regional analyses of intraspecific variability. This framework provides a robust foundation for integrating crown shape into ecological, evolutionary, silvicultural and modelling studies, including assessments of environmental effects and architectural constraints.
Maize ear geometry (length, width, curvature, and volume) is closely tied to yield and grain-filling outcomes, but existing high-throughput phenotyping pipelines remain constrained by the cost, labor, and specialized hardware they require. We developed and validated a low-cost pipeline that reconstructs a watertight 3-D mesh of a maize ear from a single 20-second video captured with a consumer-grade DSLR on a motorized turntable under uniform LED illumination. Camera poses from a multi-seed COLMAP procedure initialize a Neural Radiance Field (NeRF), and a cylindrical holder of known diameter, visible in every frame, provides automatic metric scaling with downstream geometric quality control. Applied to 300 ears spanning a diverse maize inbred panel, 250 (83.3%) passed automated processing and quality control. Skeleton length agreed with manual caliper measurements across all 250 ears (R^2 = 0.964, RMSE = 4.68 mm), and convex-hull volume agreed with water-displacement volume on a 15-ear subset spanning the full size range (R^2 = 0.982, RMSE = 5.26 mL). Residual length error grew with ear curvature, whereas bounding-box height, which records the same straight-line chord as calipers, showed no such trend; the discrepancy therefore originates in the measurement definition, since calipers record the chord while skeleton length traces the geodesic arc. The capture hardware costs approximately 607 USD, and operator involvement fell from roughly five minutes to one minute per ear, with all downstream processing running unattended. The platform provides a foundation for breeding-scale 3-D ear phenotyping.
Abstract - Plant diseases can significantly affect plant growth, productivity, and overall health. This research presents a mobile-based plant analysis system designed to identify plants, assess their health condition, and analyze visible diseases from plant images. The proposed system allows users to capture an image using a mobile camera or upload an existing image. The image is processed and analyzed using machine learning and deep learning techniques. A Convolutional Neural Network (CNN) can be used to learn visual features such as leaf shape, color, spots, and disease symptoms for plant identification and disease analysis. The system also provides a health assessment and disease severity indication to support users in understanding the condition of a plant. A Flutter-based mobile application provides the user interface, while Python and Flask can be used for image-processing and model-serving tasks. The proposed approach aims to provide a simple and accessible tool for preliminary plant identification, health assessment, and disease analysis. Key Words: plant identification, plant health assessment, disease analysis, CNN, deep learning, Flutter.
Goal. To substantiate the methodical approaches to analyzing the state of soybean crops based on the results of aerial photography with UAV by comparing manual vectorization, controlled classification according to the algorithm of maximum similarity, and uncontrolled classification of K-Means, as well as to determine the feasibility of their combination with expert visual interpretation to assess the spatial structure of the vegetation cover. Methods. Aerial photography of the test proving ground was performed with the help of the unmanned aerial vehicle DJI Phantom 4 Advanced with the subsequent photogrammetric study of materials and the formation of a highly detailed orthophotoplane. In the QGIS environment, visual decryption and manual vectorization of the main objects of the agrolandscape were carried out with the creation of polygonal layers. For automated mapping, methods of controlled classification according to the algorithm of maximum similarity and uncontrolled classification based on the K-Means algorithm were used. The accuracy of the results was evaluated by comparing the data of automated classifications with the data of manual digitization, which was used as a reference (control) method. On the basis of the results obtained, empirical data were summarized to justify practical recommendations for the application of the studied approaches. Results. The study was conducted on the territory of the research farm of the Separate subdivision of the National University of Life and Environmental Sciences of Ukraine «Berezhany Agrotechnical Institute» (vil. Pavliv, Ternopil district, Ternopil oblast) (49.452057°N; 24,805818°E) in may – august 2025. The obtained cartographic materials made it possible to quantify the areas of the main objects of the agro-landscape and identify problem areas with sparse shoots. Methods of controlled classification showed greater compliance with the digitization data (average deviation — 6.7%) compared to uncontrolled (14.7%), which were effective from the point of view of preliminary assessment of spectrally homogeneous sections, but did not provide an accurate division of shoots by density. Analysis of the spatial structure of coverage made it possible to plan local agrotechnical measures and assess the potential yield. Conclusions. Methodical approaches to analyzing the state of soybean crops based on the results of aerial photography with a UAV equipped with RGB cameras are promising and economically feasible. Automated classification methods are effective for highlighting hard and contrasting objects and small-contoured areas, while a detailed assessment of the structure of the vegetation cover is advisable to carry out using a controlled classification in combination with expert visual interpretation.
This paper presents an integrated system design for autonomous quadcopter flight path generation using the MAVLink protocol and a custom Ground Control Station (GCS) for precision agricultural crop monitoring. The system combines three coverage path algorithms (Boustrophedon, Spiral, and Energy-Optimized), a Pixhawk 4 / ArduPilot flight stack, a MicaSense RedEdge-P multispectral payload, and a ROS2-based GCS for mission planning, telemetry, and vegetation-index-based crop health assessment. The 2.8 kg quadcopter (450 mm frame, 4-cell LiPo) achieves 22–25 minutes of flight time. Across five field sizes (0.5–10 ha), the Energy-Optimized path achieved 96.5% coverage efficiency with 4.2% overlap and a 12.4% energy reduction over the Boustrophedon baseline. NDVI-based crop segmentation achieved pixel accuracy of 92.5% (maize), 94.1% (rice), and 90.8% (wheat), and four-class crop-health classification achieved a weighted F1-score of 90.0%. MAVLink 2.0 command latency averaged 15.8 ms with 99.3% packet delivery at ranges up to 800 m. An ablation study showed additional gains of 1.5–3.1% coverage from wind compensation and 2.1–2.8% from terrain-following.
Cauliflower emergence rate and seedling growth are key indicators of field conditions and varietal potential. Traditional manual surveys are unsuitable for continuous monitoring across multiple varieties. This study integrates UAV RGB imagery with the DualSlim-YOLO model to estimate cauliflower emergence rates and monitor seedling growth. Built on YOLOv11, the model incorporates a lightweight feature extraction structure and an optimized detection-scale configuration. It reduces computational complexity while maintaining detection accuracy, thereby improving the efficiency of cauliflower seedling detection. DualSlim-YOLO achieved P, R, F1-score, mAP@0.5, and mAP@0.5:0.95 of 95.35%, 96.75%, 96.05%, 98.55%, and 86.65%, respectively. The number of parameters was reduced by 38.61%, while the inference speed increased by 22.16%, demonstrating good lightweight performance. Based on this model, UAV images of 171 cauliflower varieties acquired at 7, 21, and 28 d after transplanting were used for seedling detection and emergence rate estimation. In addition, 18 time-series seedling phenotypic traits were extracted, enabling a comprehensive quantitative evaluation of emergence dynamics and early-growth performance across multiple cauliflower varieties. This method effectively screens cauliflower varieties for high emergence rates, rapid emergence, and excellent seedling growth performance. It provides technical support for high-throughput, nondestructive seedling phenotyping and early germplasm screening under field conditions.
Sebastián Varela · Jeremy Ruhter · Erik J. Sacks · Xuying Zheng · Dylan Allen · Anna Hale · Cory Landry · Xianyan Kuang · B. J. Long · Ernst Cebert · Yan Zhu · Shatabdi D. Proma · Sehijpreet Kaur · Diego Jarquin · Jesse Morrison · Andrew D. B. Leakey
Abstract The integration of digital technologies for high-throughput field phenotyping is critical for accelerating crop improvement in agriculture. However, extracting traits from remote sensing data remains constrained by fragmented workflows, manual intervention, and limited interoperability among existing tools, resulting in delays that hinder timely biological insight and decision-making. To address these challenges, we present PhenoStream (Phenotyping Streaming), a scalable, end-to-end cyberinfrastructure designed to automate the full lifecycle of aerial imagery–based phenotyping, from data acquisition to plot-and genotype-level inference. The framework integrates automated data ingestion from distributed field sites, geospatial processing, and AI-enabled trait extraction within a unified, user-accessible graphical interface. Its modular and extensible architecture supports adaptable trait modeling and seamless integration of new data sources, enabling deployment across diverse crops, environments, and experimental designs. We demonstrate the system across a large multi-location field trial network of bioenergy crops, where it enables high-throughput characterization of spatiotemporal growth dynamics, genotype-by-environment (G×E) interactions, and predictive modeling of key agronomic traits. By significantly reducing processing latency and manual effort, the platform facilitates near-real-time analysis and reproducible workflows. This work establishes a generalizable and scalable pathway for operationalizing very-high-spatial resolution aerial phenotyping in agricultural research. By bridging data acquisition and analytics, the end-to-end cyberinfrastructure provides a foundation for integrating heterogeneous and unstructured data streams—including remote sensing, environmental, and management data—toward data-driven decision making in agriculture.
Xylella fastidiosa (Xf) is among the most devastating phytosanitary threats to Mediterranean agriculture, causing Olive Quick Decline Syndrome (OQDS). Since containment efficacy depends on timely intervention, scalable in-field screening tools are needed. This study evaluates a portable digital electronic nose, based on a carbon-nanotube sensor array combined with artificial intelligence algorithms, for the in-field screening of Xf through volatile organic compound (VOC) profiling. Three replicate acquisitions were performed on 130 olive trees (390 measurements) across four cultivars (Cellina di Nardò, Ogliarola Salentina, Ogliarola Barese, and Leccino) in four Italian regions (Apulia, Calabria, Lazio, and Tuscany). Plant status was assigned from the official status of the sampling area (demarcated OQDS focus versus Xf-free area) and supported by real-time quantitative PCR (qPCR) on every plant; within demarcated sites, plants with undetectable DNA in sampled twigs were retained as Xf+ following phytosanitary criteria, giving 216 infected and 174 healthy samples. The multidimensional sensor signals were processed with an optimized Shallow Neural Network. Under plant-grouped 80/20 validation, keeping each plant’s replicates in the same subset, the model achieved (93.3 ± 4.0)% accuracy, (97.7 ± 3.5)% sensitivity, and (88.2 ± 8.3)% specificity (mean ± SD). A feature-importance analysis revealed a reproducible, though not chemically resolved, VOC-related response pattern. This low-cost, portable Internet of Things (IoT) device offers a proof-of-concept screening approach for Xf surveillance, pending plant-level and external validation.
Abstract Purpose Evaluate how soil and canopy sensing can map within-block variability in tart cherry orchards and identify indicators robust enough for repeatable management decisions. Methods Soil apparent electrical conductivity (ECa) was mapped in spring 2022 across four commercial tart cherry blocks (8.5–10.5 ha; approximately 3,500 trees per block), followed by canopy sensing in 2023–2024. Canopy structure was measured using unmanned aerial vehicle (UAV) photogrammetry and mobile terrestrial laser scanning (MTLS) using light detection and ranging (LiDAR), and canopy density using mobile ceptometry. Spatial layers were aligned to per-tree grid cells. An August 2025 campaign compared UAV- and LiDAR-derived tree height with ground-truthed height. Results Soil-to-canopy relationships were weak to moderate but consistent within blocks ( r = 0.10–0.40), with strength and direction varying by site conditions. Canopy density was more strongly associated with UAV-derived volume than height. UAV-derived 90th-percentile height best predicted ground-truthed height ( R ² = 0.89; RMSE = 0.34 m), whereas LiDAR showed a weaker relationship and greater error ( R ² = 0.70; RMSE = 0.52 m). Cross-sensor agreement was moderate to strong ( r = 0.41–0.65). Per-tree rankings were stable between years for UAV height and volume. Conclusion Whole-block sensing revealed persistent spatial patterns that could support management-zone delineation. UAV photogrammetry provided accurate canopy metrics, MTLS offered measurements suited to routine orchard operations, ceptometry added seasonal canopy-density information, and ECa provided soil context. Occasional ECa mapping combined with strategically timed UAV surveys and other sensors as needed could reduce redundant sensing while supporting fertilizer evaluation, pruning, and labor allocation.
Plant diseases are still posing a challenge to the productivity, quality of crops, and food security, especially in locations where field diagnosis is based on manual visual inspec-tion. This paper assesses deep learning network-based automated classification of plant leaf diseases on public RGB leaf-image datasets, such as the Kaggle New Plant Diseases Dataset (Augmented) and PlantVillage images. They investigated four archi-tectures: EfficientNetV2B0, ResNet152V2, DenseNet201, and one hybrid Vision Trans-former (ViT)-based model. The steps of the experiment involved loading the dataset, exploratory analysis, preprocessing, resizing, normalizing, augmentation, transfer learning, independent model training, and evaluation metrics such as accuracy, preci-sion, recall, F1-score, training curves, testing results, and confusion matrices. The hy-brid ViT-based model was reported to have the best accuracy of 99.5%. On the smaller seven class subset of PlantVillage, EfficientNetV2B0 scored 98.11%. On the 38-class dataset, DenseNet201 improved test accuracy (97.34) and validation classification ac-curacy (around 98). ResNet152V2 scored 97.01 on the 38-class test set. The results demonstrate that CNN and transformer-based models can help to recognize plant diseases accurately whereas hybrid attention-based structures provide a promising path to enhance fine-grained classification. Since the model notebooks had varying class settings and splits, the comparison is seen as a model-structured assessment as opposed to a precisely identical benchmark across all architectures.
Reproduction assets foundThe paper's phenotyping inputs are two public plant leaf-image datasets explicitly named in the Data Availability statement: the Kaggle New Plant Diseases Dataset (Augmented) and the PlantVillage dataset, both with public URLs. No author code, models, or supplementary materials are deposited (supplementary materials: 'Dataset · publicy available. Plant leaf images were
obtained from the Kaggle New Plant Diseases Dataset (Augmented) and PlantVillage
datasets. The datasets contain publicly accessible RGB images of healthy and diseased
plant leaves used for supervised image classification research.
DATASET SOURCES
Kaggle New Plant Diseases Dataset (Augmented):
https://www.kaggle.com/datasets/vipoooool/new-plant-diseases-dataset
PlantVillage Dataset: https://plantvillage.psu.edu/All processed data, experimental configurations, and model implementation details are
described within the manuscript. Additional materials may be made available from the
corresponding author upon reasonable request.
ACKNOWLEDGMENTS
The author acknowOpen asset ↗Kaggle · new-plant-diseases-datasetpdf-raw-page:24 lines:1-23Dataset · publicgmented) and PlantVillage
datasets. The datasets contain publicly accessible RGB images of healthy and diseased
plant leaves used for supervised image classification research.
DATASET SOURCES
Kaggle New Plant Diseases Dataset (Augmented):
https://www.kaggle.com/datasets/vipoooool/new-plant-diseases-dataset
PlantVillage Dataset: https://plantvillage.psu.edu/All processed data, experimental configurations, and model implementation details are
described within the manuscript. Additional materials may be made available from the
corresponding author upon reasonable request.
ACKNOWLEDGMENTS
The author acknowledges Istanbul Aydin University for academic support and research
guidance duringOpen asset ↗PlantVillagepdf-raw-page:24 lines:1-23Code / dataset availability confirmedOpenAlex · arXiv · checked 5 Sept 2026
High-throughput phenotyping depends on accurate 3D reconstruction of plants across growth stages, yet the development and evaluation of temporal completion methods are limited by the lack of datasets with complete geometric ground truth. To address this challenge, we introduce SynthCrop4D, a procedurally generated synthetic dataset of temporally evolving plant point clouds that provides controllable noise, occlusion, and complete plant geometry for benchmarking reconstruction methods. Using this dataset, we evaluate a two-stage pipeline that combines spatial denoising and temporal point cloud completion. First, a denoising module removes structural artifacts from raw laser-scanned point clouds. The resulting data are then processed by an Adaptive Temporal PoinTr model that reconstructs the current growth stage (t) using information from the previous stage (t-1), enabling recovery of regions missing due to self-occlusion. We evaluate the proposed framework on both SynthCrop4D and the real-world Pheno4D dataset (tomato and maize) under settings with and without denoising. Results show that denoising substantially improves reconstruction quality, with the best configuration achieving a Chamfer Distance of 0.0061 on SynthCrop4D (Temporal PoinTr + Mamba-DG) and an F-Score of 0.2080 on Pheno4D (Vanilla PoinTr + Mamba-DG). We further demonstrate the use of completed point clouds for phenotypic trait extraction, including plant height, canopy width, and convex hull volume, obtaining hull-volume MAEs of 0.021 on synthetic data and 0.343 on real data. Together, SynthCrop4D and the proposed pipeline provide a benchmark and methodology for temporal plant reconstruction and high-throughput crop phenotyping.
Reproduction assets foundThe paper's authors explicitly state that source code, implementation details, and pre-trained model weights are publicly available on GitHub, and that the paper-specific SynthCrop4D synthetic dataset can be reproduced via scripts in that codebase. Pheno4D is a cited prior public dataset, not a paper-specific asset.Code · publicThe source code, implementation details, and pre-trained model weights for this study are publicly available on GitHub at https://github.com/Mrudul2006/3d_plant-reconstruction .Open asset ↗Mrudul2006/3d_plant-reconstructionlines:1724-1761Plant phenotyping relevance matchOpenAlex · Europe PMC · Crossref · bioRxiv · checked 5 Sept 2026
Abstract Non-invasive, high-throughput phenotyping tools are needed that can identify environmental effects on plant structure and function to diagnose factors responsible for reduced growth in commercial and non-commercial settings. In this study, we explored whether the integration of 3D-multispectral (3D) and 2D-hyperspectral imaging (HSI), aided by machine learning (ML), could be used to identify environmental stress treatments imposed during plant growth. Controlled environment-grown Nicotiana Benthamiana plants were subjected to a range of abiotic treatments – including different growth irradiances, heat treatment and drought stress – with the treatments resulting in differences in shoot height, biomass, leaf area and spectral reflectance. ML models were trained to identify these treatments using morphological and spectral traits measured at 27, 29, 31, and 34 days after sowing (DAS). A 3D-multispectral scanner was used to obtain information on plant height, biomass, and leaf area. A visible and near-infrared (VNIR) HSI camera provided detailed spectral information for deriving spectral indices including the Normalised Difference Vegetation Index (NDVI), Photochemical Reflectance Index (PRI) and Normalized Difference Red Edge (NDRE). Manual measurements provided baseline comparative data. The 3D-multispectral scanner reliably estimated above-ground traits, with high correlations between manual and scanner-derived measurements. The ML models accurately differentiated among environmental stress treatments, with the fused 3D+HSI model achieving the best overall predictive performance across all evaluated metrics compared with models based on either imaging modality alone. Results demonstrated the effectiveness of combining 3D-multispectral and 2D-HSI data with ML analyses for non-destructive, high-throughput phenotyping. The integration of these techniques enabled non-destructive, high-throughput identification of environmental stress treatments imposed during plant growth.
Abstract Background and aims Plant volatile organic compounds (VOCs) change dynamically with plant development and in response to environmental conditions. However, their potential as non-invasive indicators of phenological progression remains poorly explored. In this study, we developed a framework integrating automated VOC sampling, time-resolved VOC profiling, and machine-learning analysis for the non-invasive assessment of plant phenology. Using soybean ( Glycine max (L.) Merr.), we investigated whether development-associated temporal variation in VOC emissions could delineate and predict developmental phases. Methods We collected VOCs daily under controlled environmental conditions from 16 to 43 days after sowing, spanning the transition from vegetative to reproductive stages, using an automated sampling system coupled with thermal desorption-gas chromatograph-mass spectrometer (TD- GC-MS). To characterise temporal changes in VOC profiles associated with phenological progression, we analysed the daily VOC data using a multi-step pipeline combining statistical filtering and similarity-based network analysis. We defined VOC-derived developmental phases from similarity patterns in the VOC profiles, then developed and evaluated machine-learning models to predict these phases. Key results Seven VOCs exhibited distinct phase-dependent dynamics, including green leaf volatiles and monoterpenes showing characteristic temporal changes during phenological progression. Network-based clustering of VOC profiles resolved five developmental phases closely aligned with conventional developmental stages. A machine-learning model predicted these phases from the VOC profiles with high predictive accuracy on independent test data, demonstrating that phenological progression could be quantitatively inferred from VOC emission patterns. Conclusions Our findings support VOC profiling as a reliable and non-invasive approach for assessing phenological progression in soybean. By extracting temporally structured VOC signals, this framework captures developmental information that may be difficult to obtain through visual observation alone, particularly after canopy closure. VOC profiling offers a practical tool for monitoring crop developmental dynamics and has broader potential for plant phenotyping and precision crop management.
Reproduction assets found保存済みの本文根拠を更新済みルールで再検証し、公開資産1件を確認しました。Supplement · publicThe peak area matrix obtained from the MS- DIAL analysis (Supplementary Dataset S1) was filtered to remove unreliable features.Open asset ↗lines:66-69Plant phenotyping relevance matchCrossref · checked 5 Sept 2026
Faba bean yield reflects complex relationships among genotype, environment, and agronomic traits. This study evaluated an explainable Tabular Prior-data Fitted Network (TabPFN) framework for estimating plot-mean single-plant yield and prioritizing traits using 398 plot-level observations, 13 measured agronomic predictors, and six derived features. On the reference 80/20 split, TabPFN achieved the best values for all four test metrics (R2 = 0.8746, RMSE = 1.9132 g plant−1, MAE = 1.0819 g plant−1, and MAPE = 8.16%). The Friedman test detected differences among the six models (χ2(5) = 16.75, p = 0.005); Nemenyi comparisons distinguished TabPFN from HistGradientBoosting and SVR, whereas the Holm-corrected Wilcoxon analysis confirmed only the TabPFN–SVR difference. Across 10 repeated 80/20 splits, TabPFN obtained the highest mean test R2 (0.8614 ± 0.0691), ranked first in eight splits, and produced a higher R2 than every tuned baseline in at least eight splits. SHAP, permutation importance, and LOCO analyses emphasized pod-, seed-, and biomass-related predictors. Repeated-split ablation showed that derived features improved TabPFN consistently, whereas removing selected target-proximal yield variables reduced performance for every model. The framework is therefore a harvest-time trait-estimation and trait-prioritization tool rather than an early-season forecasting system. Notably, TabPFN achieved this performance without the 100-trial Optuna search used for each baseline; only n_estimators was screened over four prespecified values.
While most phenotyping platforms rely primarily on image-based measurements, advanced plant characterization requires the integration of active physiological sensing modali- ties such as chlorophyll fluorescence. We present an autonomous robotic platform designed to perform targeted fluorescence measurements on plant leaves. The system combines 3D plant reconstruction, geometric analysis, and motion planning to localize suitable measurement points and generate collision-free trajectories for a robotic manipulator. A dense 3D model of the plant is reconstructed from multi-view data and used to extract candidate leaf surfaces based on orientation, accessibility, and sensing constraints. These targets are then integrated into a task-level planning framework that guides the end-effector to precise contact or near-contact configurations required for point-based fluorescence acquisition. The platform enables automated, repeatable, and spatially resolved physiological measurements that go beyond passive imaging. By tightly coupling perception, geometric reasoning, and manipulation, the proposed system provides a robotics-driven approach to high-resolution plant phenotyping and opens new directions for autonomous agricultural inspection and plant-aware manipulation.
Reproduction assets foundThe paper states its code is publicly available in the authors' SonyCSLParis GitHub repository (Plant3DImager), which implements the phenotyping perception and motion-planning pipeline. The exact full URL is split across a line break in the supplied text, so the verifiable allowed URL prefix is used.Code · public2 The code is available at https://github.com/SonyCSLParis/ 3 See for example at https://www.youtube.com/watch?v=Open asset ↗SonyCSLParis/pdf-page:4 lines:1-61Plant phenotyping relevance matchOpenAlex · checked 11 Sept 2026
Accurate quantification of rice aboveground biomass (AGB) is critical for crop monitoring but remains challenging due to the complex nonlinearity arising from the coupling of plant density, spatial structure, and internal dry matter distribution. To address the limitations of single-source remote sensing, this study proposed a Three-Dimensional Dry Matter Distribution Integration (3D-DMI) model, which establishes a physically interpretable framework decomposing AGB into dry matter density ( ρ ), horizontal projection distribution ( S ), and vertical cumulative distribution ( h d ) components. Guided by this framework, a core subset of six features (Red_650, MTCI, G_correlation, R_correlation, LPI, and HPA0_99) was extracted from UAV-based multispectral, RGB, and LiDAR data using a dual-step feature selection approach combining Maximum Information Coefficient (MIC) and Distance Correlation (dCor). A Random Forest (RF) regression model was then developed to estimate AGB across the entire growth season. The results demonstrated that the 3D-DMI model achieved excellent performance with an R 2 of 0.920, an RMSE of 0.184 kg/m², and an RPD of 3.544, significantly outperforming any single-sensor approach. Single-feature analysis revealed that while LiDAR-derived structural features provided the fundamental basis for biomass estimation, they encountered inherent saturation bottlenecks during late growth stages. Feature contribution analysis based on SHAP further quantified that LiDAR-derived features dominated the estimation process (68.5% contribution), providing the volumetric basis, whereas RGB textures (18.3%) and multispectral features (13.3%) provided indispensable supplements. Ultimately, this study established a robust, physically grounded computational paradigm for high-precision UAV-based rice biomass monitoring across the entire growth cycle.
Sebastian Varela · Jeremy Ruhter · Erik Sacks · Xu-ying Zheng · Dylan Allen · Anna Hale · Cory Landry · Xian-yan Kuang · Blake Long · Ernst Cebert · Yan Zhu · Shatabdi D. Proma · Sehijpreet Kaur · Diego Jarquin · Jesse Morrison · Andrew D. B. Leakey
Abstract The integration of digital technologies for high-throughput field phenotyping is critical for accelerating crop improvement in agriculture. However, extracting traits from remote sensing data remains constrained by fragmented workflows, manual intervention, and limited interoperability among existing tools, resulting in delays that hinder timely biological insight and decision-making. To address these challenges, we present PhenoStream (Phenotyping Streaming), a scalable, end-to-end cyberinfrastructure designed to automate the full lifecycle of aerial imagery-based phenotyping, from data acquisition to plot- and genotype-level inference. The framework integrates automated data ingestion from distributed field sites, geospatial processing, and AI-enabled trait extraction within a unified, user-accessible graphical interface. Its modular and extensible architecture supports adaptable trait modeling and seamless integration of new data sources, enabling deployment across diverse crops, environments, and experimental designs. We demonstrate the system across a large multi-location field trial network of bioenergy crops, where it enables high-throughput characterization of spatiotemporal growth dynamics, genotype-by-environment (G×E) interactions, and predictive modeling of key agronomic traits. By significantly reducing processing latency and manual effort, the platform facilitates near-real-time analysis and reproducible workflows. This work establishes a generalizable and scalable pathway for operationalizing very-high-spatial resolution aerial phenotyping in agricultural research. By bridging data acquisition and analytics, the end-to-end cyberinfrastructure provides a foundation for integrating heterogeneous and unstructured data streams-including remote sensing, environmental, and management data - toward data-driven decision making in agriculture.
Chlorophyll fluorescence provides sensitive information on plant photochemical responses, but its measurement requirements can limit high-throughput application. This study investigated whether RGB imagery could approximate chlorophyll-fluorescence-derived photochemical status across garden plant species during progressive soil drying. A Photochemical Status Index (PSI) was constructed by principal component analysis from five highly correlated JIP-test energy-flux variables (RC/CS, ABS/CS, TRo/CS, ET2o/CS, and RE1o/CS). The dataset comprised 50 aggregated species-by-soil-moisture-stage observations representing ten species and five sequential soil-moisture stages. Eleven RGB-derived variables were evaluated, and a partial least-squares regression model was assessed using nested leave-one-species-out validation, with all data-dependent procedures repeated within each outer training fold. PC1 explained 96.5% of the shared variation among the fluorescence-derived fluxes. The predictors g, GLI, ExG, ExGR, and CIVE were retained in all ten outer folds. The final model yielded a pooled out-of-fold R2 of 0.469, an RMSE of 1.585, and an MAE of 1.183. However, species-specific R2 ranged from −0.179 to 0.959, and a calibration slope of 0.509 indicated prediction-range compression. These findings provide proof-of-concept evidence of moderate RGB-based approximation of fluorescence-derived photochemical status, but inconsistent species transferability and the common soil-moisture/time gradient require external validation before practical deployment.
Agricultural operations such as pepper harvesting, fruit counting, and field phenotyping rely on accurate visual recognition and instance segmentation algorithms. However, pepper fruits in complex field environments often exhibit slender and curved shapes, partial occlusion, ambiguous boundaries, and adhesion between adjacent instances. Existing object detection and instance segmentation methods therefore struggle to obtain complete fruit masks, which adversely affects subsequent fruit counting, contour measurement, and picking-point localization. To improve the instance segmentation accuracy of occluded peppers in complex field scenes, this study proposes OccPepSeg-YOLO, an improved model based on YOLO11n-seg. First, a P2FreqFusion module is introduced to fuse shallow, high-resolution detail features with deep semantic features, thereby enhancing the representation of fruit edges and tip regions. Second, an ASC module is designed to model the directional and scale-related morphological characteristics of pepper fruits, while a BoundaryGate module strengthens responses at occlusion interfaces and boundaries between adjacent instances. Finally, an OccPepSegment multi-scale prototype segmentation head is constructed, and a BDoU loss function is introduced to improve the boundary consistency of instance masks. Experiments on a self-constructed field-pepper instance segmentation dataset showed that OccPepSeg-YOLO achieved M-P, M-R, M-mAP50, and M-mAP50–95 values of 93.87%, 92.09%, 97.17%, and 82.31%, respectively, representing improvements of 5.59, 3.18, 3.83, and 9.52 percentage points over YOLO11n-seg. Further comparisons with representative YOLO-based instance segmentation models, including YOLOv8n-seg, YOLOv9c-seg, YOLO12n-seg, and YOLOv26n-seg, demonstrated that OccPepSeg-YOLO achieved the best overall segmentation performance. In particular, its M-mAP50–95 exceeded the best competing result obtained by YOLOv9c-seg by 8.35 percentage points. Under a unified repeated-inference protocol on an RTX 3090 GPU using FP32 precision, a batch size of 1, and 640 × 640 inputs, OccPepSeg-YOLO achieved a mean inference latency of 15.801 ± 1.238 ms, a P95 latency of 17.323 ms, and a throughput of 63.29 FPS. These results demonstrate that the proposed model can produce more complete pepper instance masks under leaf occlusion, fruit overlap, and complex background conditions, providing technical support for field-pepper recognition, fruit counting, and visual perception by agricultural robots.
High-resolution monitoring of forest structure and productivity is essential for effective natural resource management. However, monitoring approaches such as field-based forest inventories or extensive lidar campaigns are costly, time-intensive, and spatially limited. Therefore, inexpensive and accessible methods are needed. SatCHM (Satellite Canopy Height Model) was developed to be an accessible and open-source tool for researchers, allowing for site-specific and temporally flexible predictions of canopy height with limited computational resources. SatCHM requires four inputs: panchromatic satellite imagery, solar and sensor angle metadata of satellite imagery, digital elevation models (DEMs), and lidar-produced CHMs for an area of interest. After SatCHM pre-processes inputs, data is loaded into a collection of convolutional neural networks (CNNs) for image-to-image regression. This ensemble cooperates to yield high-resolution predictions (up to 0.5-meter) of three-dimensional tree structure with discernible tree crowns across a broader defined area of interest. After calculating the mean absolute error for each prediction output, the median of these mean absolute errors was 6.06 meters.
Lodging in sorghum presents a significant challenge for plant breeders due to the trade-off between lodging resistance and grain yield. Manually measuring lodging across thousands of plots is time-consuming, expensive, and error-prone, making selection for lodging resistance challenging in breeding programs. Unmanned aerial vehicle (UAV)-derived metrics provide a potential high-throughput alternative; however, it remains unclear whether photogrammetric heights derived from UAV imagery can estimate plot-level lodging severity in large sorghum breeding trials. This study developed a framework for predicting plot-level lodging from UAV imagery across 2,675 sorghum breeding plots. Multi-temporal canopy height data were collected at two critical time points: maximum crop height and at manual lodging assessment. Height percentiles were extracted from UAV-derived point clouds generated using photogrammetric algorithms. These data were used to develop parametric, non-parametric, and ensemble prediction models, which were evaluated using three statistical metrics. The ensemble model, averaging predictions from all models, achieved the highest accuracy with Pearson correlations of r = 0.80-0.84 and lowest root mean square error (RMSE=16-18%), explaining 64-70% of variation in manual lodging counts. Model diagnostics and iterative refinement, including inspection of UAV imagery and dataset curation, had minimal impact on model performance, demonstrating the robustness of the approach. Model performance was consistent across sites, with minimal effects of stratified sampling on accuracy, confirming the ensemble approach as optimal for plot-level lodging assessment. This study demonstrates that integrated multi-temporal UAV imagery offers a practical alternative to labor-intensive manual evaluation methods by enabling high-throughput lodging assessment suitable for implementation in sorghum breeding programs.
Abstract Deep convolutional networks now classify leaf images on curated benchmarks such as PlantVillage with accuracies close to the measurement ceiling of those datasets, which has shifted the open questions away from raw accuracy toward two under-reported issues: which components of a composite pipeline actually cause the result, and whether the components that matter on laboratory images are the same ones that matter on field photographs. We address both with a classification framework evaluated under a protocol that fixes every development decision before the independent test data are read. A Mask R-CNN stage localizes the dominant leaf, the accepted box is expanded by a validation-selected 5% margin and resized to a shared 224 by 224 input, and the same localized image is passed to ResNet50, InceptionV3, and MobileNetV2 together with an extractor that produces eighteen colour and shape descriptors. The forty-five class probabilities and eighteen descriptors form a sixty-three-dimensional input to a neural meta-classifier developed by three repetitions of stratified five-fold cross-validation. On a locked 3,101-image PlantVillage test partition of fifteen classes the framework reached 99.77% accuracy and 99.75% macro F1 with seven misclassifications, and a one-component-at-a-time ablation confirmed that every stage contributed. The same design was then trained and evaluated entirely within a separate thirteen-class PlantDoc field-image dataset, where it reached 90.03% accuracy and 89.40% macro F1. This is a within-PlantDoc experiment and not a controlled-to-field transfer test, so the figure measures how the pipeline behaves on field imagery rather than how a PlantVillage-trained model survives a domain shift. The central finding comes from running the identical ablation on both datasets: on clean images the ensemble breadth and descriptors provide the incremental gains, but under field conditions the ordering changes, and leaf localization and learned fusion become the decisive components. Removing localization cost 2.83 accuracy points and replacing the neural fusion with soft voting cost a further 2.08 points, the two largest effects on PlantDoc. The contribution is a controlled and transparent account of where each component of a localization-aware plant disease classifier earns its place, and of how that ordering shifts between laboratory and field acquisition.
Reproduction assets found保存済みの本文根拠を更新済みルールで再検証し、公開資産1件を確認しました。Dataset · publicThe classification subset used here (20,638 images spanning fifteen pepper-bell, potato, and tomato classes) was obtained from PlantVillage, which is openly accessible at https://www.kaggle.com/datasets/emmarex/plantdisease.Open asset ↗Kaggle · emmarex/plantdiseaselines:314-336Plant phenotyping relevance matchCrossref · checked 5 Sept 2026
Introduction Accurate estimation of aboveground biomass (AGB) is essential for monitoring pasture productivity and supporting sustainable management of integrated crop–livestock (ICL) systems. We hypothesized that integrating multispectral, thermal, and canopy-structural information derived from unmanned aerial vehicles (UAVs) would improve AGB prediction relative to spectral information alone, and that Generalized Additive Models for Location, Scale and Shape (GAMLSS) would accommodate seasonal heteroscedasticity while maintaining predictive performance comparable to Random Forest (RF) and Support Vector Machine (SVM) models. Methods We collected 280 destructive biomass samples from two ICL paddocks and one continuously grazed pasture in the Brazilian Cerrado between 2022 and 2024. Twenty-four UAV-derived predictors, including spectral bands, vegetation indices, canopy surface temperature, and canopy height, were evaluated using repeated five-fold cross-validation. Model transferability was assessed by withholding one management paddock at a time. Results and discussion Under repeated five-fold cross-validation, GAMLSS achieved the lowest prediction error (R² = 0.69 ± 0.01; RMSE = 2.15 ± 0.04 Mg ha⁻¹), followed closely by SVM (R² = 0.68 ± 0.01; RMSE = 2.19 ± 0.03 Mg ha -1 ); RF showed lower accuracy (R 2 = 0.53 ± 0.01; RMSE = 2.63 ± 0.02 Mg ha -1 ). In the paddock-transferability assessment, GAMLSS also showed the lowest error (R 2 = 0.63 ± 0.04; RMSE = 2.34 ± 0.26 Mg ha -1 ). For GAMLSS, the complete multisensor configuration reduced RMSE by 6.2% compared with the spectral-only configuration. The selected model was used to generate spatially explicit maps of AGB and standing aboveground biomass carbon, estimated from the mean measured carbon concentration of forage biomass. Integrating multispectral, thermal, and structural UAV data with distributional regression improves AGB estimation and enables spatial monitoring of tropical pastures under contrasting management conditions.
Ilya A. Grishin · Gennady I. Afanasyev · Valeri I. Terekhov
Field / plotLiDAR / point cloudStem / branchWhole plant / canopy / plot / fieldClassification2D/3D reconstructionSegmentationArchitecture / morphology / geometry
The separation of a forest plot into individual trees and the automatic extraction of their stem structures from terrestrial laser scanning data are complicated by dense stands, occlusions, and the diversity of biomorphological forms. Existing algorithms usually solve scene partitioning, voxel classification, and tree growing as independent tasks, which leads to error accumulation at subsequent processing stages. This paper proposes a unified model for spatial fragmentation and biomorphological forest segmentation comprising three interrelated stages: scene partitioning by estimated stem coordinates using a Voronoi diagram, probabilistic voxel- or point-level classification, and bottom-up tree growing guided by spatial connectivity and stem membership criteria. For the semantic module, tabular, volumetric, and point-based approaches are compared: gradient boosting with layer-by-layer inference, TabNet, a three-dimensional convolutional neural network, PointNet2, and two-stage pipelines in which gradient boosting builds an initial stem mask for subsequent neural segmentation. The experiment was conducted on 567 mixed-species trees. Considering both quality and computational performance, the {CatBoost; CNN3D} pipeline was selected as the preferred solution, achieving AUC = 0.9966 and IoU = 0.9831. The obtained results show that combining interpretable layer-by-layer classification with subsequent spatial analysis improves the quality of stem structure reconstruction, which is important for automatic forest inventory tasks.
To establish an accurate and interpretable prediction framework for soybean lodging grade and clarify the core regulatory traits and differentiated driving mechanisms of soybean lodging under high-density drip irrigation cultivation, 356 spring soybean germplasm accessions were used as experimental materials in this study. Morphological and mechanical traits including plant height (PH), stem pulling force (SPF), internode number (IN) and petiole length (PL) were measured over two consecutive years of field phenotyping. Two composite evaluation indices, plant height/stem pulling force ratio (PH/SPF) and plant height/internode number ratio (PH/IN), were further constructed. Four machine learning algorithms were adopted to develop multi-classification models for soybean lodging grade prediction. SHAP analysis combined with three global sensitivity approaches (perturbation analysis, Sobol’ method and Morris screening) was applied to decipher the regulatory patterns of key traits. The results showed that lodging grade significantly affected soybean grain yield and explained 25–28% of the phenotypic yield variation; yield reduction tended to plateau under severe lodging. Compared with single indicators such as SPF and PL, the two derived composite indices could stably distinguish soybean accessions with different lodging grades and exhibited stronger discriminatory power. Model comparison revealed that the XGBoost model achieved optimal prediction accuracy and generalization stability for lodging grade, with a weighted F1-score of 95.34% on the test set, significantly outperforming the conventional linear model. Interpretability analysis demonstrated that the PH/IN, PH, and PH/SPF acted as the primary positive traits promoting lodging, while SPF was the sole protective trait. Driving factors of lodging presented obvious gradient heterogeneity: mild lodging was dominated by the imbalance of plant architecture ratio, whereas severe lodging was governed by the cumulative effects of PH and IN. Strong interactions existed among all measured traits. The interpretable machine learning framework established in this study can provide theoretical support and technical references for lodging-resistant germplasm screening and targeted plant architecture regulation for densely planted soybean under drip irrigation systems.
Microplastics (MPs) are persistent and ubiquitous contaminants in aquatic ecosystems, yet their interactions with submerged aquatic plants remain poorly understood. While MP-induced phytotoxicity has been extensively investigated in terrestrial plants, quantitative evidence for MP uptake and internal accumulation in submerged species is still limited. In this study, we investigated the phytotoxicity and accumulation patterns of fluorescent microplastics (FMPs) in two submerged aquatic plants, Bacopa lanigera and Rotala indica, using fluorescence spectroscopy. Plants were exposed to FMPs of two particle sizes (50 nm and 1 µm) across three exposure concentrations (0.001%, 0.01%, and 0.05%). Plant growth, chlorophyll content, fluorescence emission, and FMP accumulation were systematically evaluated. Our results demonstrated clear size- and concentration-dependent responses. Smaller particles (50 nm) showed significantly higher uptake and induced stronger phytotoxic effects than 1 µm particles, with pronounced growth inhibition and chlorophyll reduction observed at the highest concentration (0.05%). Fluorescence-based analysis enabled quantitative estimation of both surface-associated and internalized FMPs within plant tissues. Maximum surface accumulation reached 207 ppm, while internal (cross-sectional) accumulation reached up to 75 ppm, regardless of plant species. Under the respective experimental conditions, B. lanigera exhibited higher estimated FMP accumulation, whereas R. indica showed greater growth inhibition. These findings provide quantitative evidence of microplastic uptake and internal accumulation in submerged aquatic plants and highlight particle size as a critical determinant of phytotoxicity. Moreover, this study establishes a fluorescence-based methodological framework for estimating microplastic concentrations in aquatic plant tissues, contributing to improved ecological risk assessment of microplastics in freshwater ecosystems.
Abstract Purpose : The system for diagnosing diseases in Solanaceae crops (SolanAPP), including tomatoes, potatoes, peppers, and eggplants, represents a promising tool for supporting decision-making in agricultural fields using AI. This system reconciles the computational intensity of multitasking models with the infrastructural limitations of rural environments, thereby increasing digital literacy. Its architecture is based on two fundamental pillars: (i) autonomous, offline operation for the detection and classification of diseases in Solanaceae crops; (ii) a georeferenced epidemiological surveillance network with agricultural recommendations for crop monitoring. Methods : The core diagnostic process combines crop-specific semantic segmentation and disease classification models exported to TensorFlow Lite, enabling on-device visual inference and pixel-level severity estimation. An optional online layer integrates Groq’s large language model (LLM)-based reasoning and Firebase services to generate structured agronomic explanations and facilitate the creation of georeferenced community reports when a connection is available. Preprocessing and management of the dataset were performed using the Roboflow platform. The mobile app was developed natively in Kotlin. Model performance was rigorously evaluated using accuracy, recall, F1 score, mean IoU, mPA, inference latency, model size, and decision matrix. The optimal model for the crop was selected using Simple Additive Weighting (SAW). Finally, the overall framework quality and usability were evaluated in Cuba through a user validation survey using a 5-point Likert scale and aligned with the ISO/IEC 25010 software quality model. Results : The model that yielded the best results for most crops was DeepLabV3+ with MobileNetV2, which achieved a classification accuracy of over 97\% while operating with lower inference latency. Beyond individual diagnoses, the system incorporates a collaborative georeferencing feature that allows users to share observations and precise geographic coordinates of detected pathologies to facilitate regional epidemiological monitoring. The user satisfaction survey yielded a satisfaction rating of 4.5/5, with users highlighting the importance of offline diagnosis. Conclusion : Plant disease diagnosis using computer vision can support earlier intervention in resource-constrained agricultural settings, but practical deployment requires models that are accurate, lightweight, interpretable, and usable under limited connectivity. SolanAPP, an offline-first Android framework for detecting foliar pathologies in Solanaceae crops, not only establishes a framework for disease identification in complex natural environments but also provides a theoretical and practical foundation for automated agronomic treatment recommendations and community-based crop surveillance. Impact SolanAPP is a free framework that supports the synergy between multitask deep learning for offline disease diagnosis and LLM-driven reasoning for decision-making in the field. Beyond the quantitative metrics obtained from the selected models, the deployment of SolanAPP in rural contexts serves a fundamental socio-technical purpose: it acts as a catalyst for open access, digital literacy, and agronomic decision-making under unfavorable development conditions. It also represents a strong effort to foster a collaborative epidemiological surveillance network in the agricultural sector. Although it faces challenges, such as the use of field images for model training, this framework marks a promising step in the deployment of edge AI, balancing technical accuracy with practical utility.
Abstract Plant diseases represent major constraints on agricultural productivity, often resulting in significant yield losses. This study presents FitoView, a cloud-based mobile decision support system that integrates deep learning with real-time weather forecasting for sustainable plant disease management. Demonstrated through a case study on Cercospora leaf spot in chili pepper, the system employs custom YOLOv8 models for automated disease detection, classification, and pixel-level severity quantification, combined with meteorological data from the OpenMeteo API. The core innovation lies in an integrated decision matrix that considers three dimensions: AI-assessed disease severity, 48-hour climatic risk forecasts, and optimal spraying conditions, generating four contextualized management scenarios with tailored re-evaluation periods (3-10 days). OpenMeteo API validation across four cities in Sergipe demonstrated very strong correlations, with Pearson coefficients (r) of 0.90-0.97 for temperature, 0.81-0.95 for humidity, and 0.92-0.95 for solar radiation, corresponding to R² values of 0.65-0.94. The YOLOv8 object detection model achieved perfect precision (100%) and macro-averaged recall of 89% across all disease classes, with Cercospora leaf spot detection reaching perfect metrics (100% precision, recall, and F1-score). Field validation in Lagarto, Sergipe, confirmed the system’s practical use: it accurately detected Cercospora leaf spot, estimated severity, and, combined with climatic risk, generated recommendations for alternative treatment and short-term re-evaluation. The Progressive Web Application architecture, deployed on a Cloud Platform, ensures accessibility without installation requirements, while the modular design enables scalability to additional crops and diseases, representing a significant advancement toward democratizing AI-powered precision agriculture tools for smallholder farmers in Brazil.
Maize ( Zea mays L.) is a highly adaptable crop grown worldwide across diverse climates and management practices, with uses across multiple sectors. Consequently, the traits prioritized in maize research and breeding programs vary depending on the specific objectives. Core traits, however, such as flowering time, plant and ear height, and stalk and root lodging, which are important for evaluating and improving the performance and stability of maize genotypes, are routinely evaluated across breeding programs, regardless of their goals. Standardized measurement of these core traits is essential to ensure data reliability and comparability, enabling the integration of phenotypic data across different experiments. Such efforts ultimately support better decision-making and accelerate the development of improved maize genotypes. This is particularly important in public sector programs, where large-scale evaluations, critical for assessing the value of specific genotypes, are often only feasible through collaboration across programs. Here, we provide a protocol for the standardized collection of phenotypic data, specifically focusing on how to measure core traits in maize field trials. These methods promote consistency and accuracy in the evaluation of these traits, and support communication and coordination among groups in the public sector and other research settings. Further, such standardization facilitates the integration and comparison of data across programs, enabling robust longitudinal and multienvironment analyses.
Florian Tanner · Judith Atieno · Sara N. Blake · M. Krysinska-Kaczmarek · Chris Brien · Mohsen Khani · K.R. Clarke · Darren Plett · Jennifer Davidson · Bettina Berger
Abstract Ascochyta blight is a widely occurring chickpea fungal disease that can cause severe yield loss. Breeding for crop resistance benefits from high‐throughput evaluation of plant–pathogen interactions in genotypes which can serve as sources of resistance. Current practice for the evaluation is human visual scoring of disease symptoms, which is limited in throughput and precision. Here, we developed open‐source sensor‐based phenotyping methods using red, green, blue (RGB) and multispectral imaging to measure resistance components and predict disease severity classes in chickpea and wild relatives grown outdoors over three seasons. Pots were imaged at multiple time points with a ground‐based platform, providing 86,792 RGB and 8199 multispectral images. Lesion count was estimated with YOLOv5 (You Only Look Once version 5) object detection (F1 score = 0.27–0.30), fractional green canopy cover was estimated from RGB images, and vegetation indices were extracted from multispectral images. A model trained on growth rates of fractional green canopy cover normalized to control genotypes could predict disease severity classes with an accuracy of 65% –81 % ( 0.43–0.59) on unseen data from three different seasons. The developed methods provide a pathway to predict visual disease severity scores and support the breeding of crops for disease resistance. They may also be used to characterize disease progression, to find underlying resistance mechanisms, and for early disease detection.
Abstract This study evaluated the potential of UAV-derived RGB spectral indices to predict fiber quality traits in cotton genotypes. Nineteen genotypes were assessed under a randomized complete block design, and RGB imagery acquired at full flowering was used to calculate GLI, NGRDI, SCId, and SI. Significant genetic variability and moderate-to-high heritability were observed for both fiber traits and spectral indices. GLI was positively associated with the Spinning Consistency Index, whereas NGRDI was associated with fiber length uniformity. Regression models showed moderate predictive ability (R² LOOCV between 32.4–35.4%; and accuracy between 0.57–0.60). GLI and NGRDI demonstrated potential as complementary tools for large-scale phenotyping and preliminary genotype selection, although they do not replace conventional fiber quality analyses. Further studies across additional developmental stages are needed to improve prediction accuracy.
The rapid detection and continuous monitoring of crop health and disease outbreaks are critical components of modern precision agriculture, essential for maintaining global food security. Traditional field-based scouting methods, while accurate, are often labor-intensive, time-consuming, and limited by spatial coverage, making them inadequate for large-scale agricultural operations. Remote sensing (RS) technologies—spanning satellite imagery, drone-based aerial platforms, and proximal sensors—offer a powerful, non-destructive, and scalable alternative for capturing high-resolution spectral and temporal data. This paper provides a comprehensive evaluation of current remote sensing applications in crop health monitoring and disease surveillance. We analyze how vegetation indices derived from multispectral and hyperspectral data, such as NDVI and red-edge parameters, serve as sensitive indicators of physiological stress and pathogen infection, often manifesting before visible symptoms appear. Furthermore, we explore the integration of machine learning and artificial intelligence algorithms in automating disease identification and severity mapping. By synthesizing recent advancements in sensor technology and data analytics, this paper demonstrates that remote sensing is indispensable for proactive, site-specific management. The findings emphasize that a multi-scale RS approach—integrating broad-scale satellite monitoring with high-resolution drone sorties—enables farmers to optimize input efficiency, minimize yield losses, and enhance the overall resilience of agro-ecosystems against biotic and abiotic stressors.
Plant disease and plant stress early warning systems have advanced through deep learning, remote sensing, digital phenotyping, disease forecasting, and sensor networks. Detection accuracy, precision, recall, F1-score, and area under the curve remain indispensable, but they are insufficient for judging whether a warning can support timely and proportionate phytoprotection under field variability. This Mini Review argues that intelligent plant health warning systems should be evaluated not only as prediction models, but also as safety-relevant decision-support systems embedded in biological, agronomic, and operational contexts. We first relate AI-based detection to established plant disease forecasting and decision-support traditions, including weather-based models, epidemiological forecasting, and integrated disease management. We then adapt selected safety-assurance concepts, including risk assessment, failure mode and effects analysis, Bow-tie reasoning, warning-threshold governance, reliability analysis, resilience thinking, and response closure, to host-pathogen-environment warning chains. The proposed framework links AI or sensor outputs with pathogen biology, host susceptibility, environmental conduciveness, inoculum pressure, uncertainty assessment, risk classification, threshold decisions, human or automated verification, intervention, and feedback learning. Illustrative crop-pathogen scenarios, including wheat rust, rice blast, potato late blight, grapevine downy mildew, and citrus greening, show how safety assurance can complement existing forecasting and decision-support systems rather than replace them. The framework remains conceptual, and whether these added assurance functions improve existing warning systems requires comparative evaluation under field conditions. Future systems should be evaluated through detection performance and response-oriented indicators such as lead time, calibration, false-alert burden, missed-warning rate, response completion, disease suppression, economic value, and learning after field action.
Precision agriculture increasingly requires intelligent systems capable of integrating multimodal sensing with transparent decision support to enable timely and reliable crop management. This study proposes a hybrid intelligent IoT framework integrating environmental monitoring, wearable plant physiological sensing, AI-based pest-monitoring, machine-learning-based prediction of crop physiological stress, and explainable fuzzy rule-based decision support into a unified architecture for crop stress assessment. A novel Physiological Stress Index (PSI) was developed by combining vapor pressure deficit, relative humidity, Delta-T, leaf capacitance, and relative irradiance to provide an interpretable indicator of crop physiological stress. The proposed framework was experimentally validated under real field conditions using a commercial environmental monitoring station, wearable leaf sensors, AI-enabled pest-monitoring devices, and cloud-based analytics. Correlation analysis confirmed strong relationships between PSI and the principal environmental variables (VPD: r = 0.980, Delta-T: r = 0.990, RH: r = −0.961), demonstrating the internal consistency and sensitivity of the proposed index. At the 15 min forecasting horizon, Linear Regression and Gradient Boosting demonstrated virtually identical performance: Gradient Boosting achieved a marginally lower RMSE and higher R2 (RMSE = 0.0273; R2 = 0.9810), whereas Linear Regression achieved a slightly lower MAE (MAE = 0.0186). At the 1 h forecasting horizon, Gradient Boosting achieved the strongest performance (R2 = 0.9034), indicating increasing relevance of nonlinear modelling at longer prediction horizons. The proposed framework demonstrates the feasibility of combining multimodal sensing, machine learning, explainable artificial intelligence, and edge-enabled IoT technologies to support proactive, transparent, and intelligent precision agriculture.
Abstract. In the face of escalating global food demand and increasing climate variability, precise and granular crop yield monitoring is indispensable for maintaining regional agricultural stability. However, current deep learning approaches for yield estimation are severely constrained by their heavy reliance on massive in situ labeled data, which limits their application in data-scarce regions. Furthermore, these models often overlook the essential temporal evolution logic of yield formation and lack a systematic discussion regarding the contribution patterns of different feature dimensions, resulting in a black-box nature of the underlying model mechanisms. To address these challenges, this study proposes a field-label-free training framework for maize yield estimation that couples mechanistic model with deep learning. The framework's core strength lies in a physiologically complete simulation database, using the WOFOST model to exhaustively cover 30 years of climate variability and habitat combinations across Northeast China (1.24 × 106 km2). A Gated Recurrent Unit (GRU) network was then introduced for end-to-end modeling, accurately capturing the energy accumulation trajectory from vegetative to reproductive growth. Validation against 458 independent ground points (2022–2024) demonstrated robust generalization with an R2 of 0.69, an RMSE of 1.21 t ha−1, and an RRMSE of 13.73 %, despite using no ground data for training. Our analysis revealed that integrating photosynthetic intensity (LAImean), duration (LAD) and peak features (LAImax) across growth stages is critical for accuracy, while omitting early-stage features significantly impairs the model's ability to capture cumulative growth effects. Furthermore, the model successfully captured the spatiotemporal yield anomalies caused by the 2023 typhoon and flooding events. Ultimately, this study generated a 10 m resolution maize yield dataset (2019–2024) for Northeast China. The dataset exhibits consistent interannual stability, with the RRMSE ranging from 7.98 % to 12.92 % and the R2 remaining above 0.44 at the city level. By deeply coupling mechanistic simulation with data mining, this dataset provides detailed support for optimizing agricultural production and guiding farming practices. The Northeast China Maize Yield 10 m dataset is openly available at https://doi.org/10.5281/zenodo.19547014 (Hu et al., 2026).
Reproduction assets foundThe paper's core output, the NortheastChinaMaizeYield10m maize yield dataset (2019–2024) with accompanying uncertainty layers, is openly deposited on Zenodo with an explicit availability statement and DOI. No author analysis code or trained model checkpoints are stated as publicly available.Dataset · publicThe Northeast China Maize Yield 10 m dataset is openly available at https://doi.org/10.5281/zenodo.19547014 (Hu et al., 2026).Open asset ↗Zenodo · 10.5281/zenodo.19547014lines:158-191Plant phenotyping relevance matchCrossref · checked 5 Sept 2026
Plant diseases destroy 20–40% of global food production annually, posing a critical threat to food security for a projected population of 9.7 billion by 2050. Conventional diagnostic approaches relying on expert visual assessment are slow, costly, and unsuitable for modern agricultural scales. While deep convolutional neural networks demonstrated early promise, single-modality, image-centric systems consistently fail under real-world field conditions characterized by variable lighting, co-occurring infections, and cultivar diversity. This review synthesizes a decade of progress across four interconnected frontiers: the evolution of deep learning architectures for plant disease detection; the adaptation of foundation models including CLIP, SAM, and DINOv2 to agricultural contexts; the development of multimodal fusion frameworks integrating imagery, environmental, genomic, and hyperspectral data; and the transition from static disease diagnosis to descriptive comparison of reported metrics, which suggested that multimodal approaches frequently reported improved diagnostic performance relative to corresponding single-modality baselines, although direct cross-study comparison was limited by methodological heterogeneity. A systematic review following PRISMA guidelines identifies eligible comparative studies. Descriptive comparison of reported performance metrics across these studies indicated that multimodal approaches generally achieved higher accuracy and sensitivity than single-modality models, particularly for pre-symptomatic disease detection. Eight critical research gaps are identified, including the absence of a unified agricultural foundation model and limited climate-aware forecasting under non-stationary climate projections. A structured research agenda is proposed to accelerate translation from laboratory performance to globally equitable, field-deployable crop protection systems.
Abstract In modern agriculture, it is essential to identify the early symptoms of plant diseases and to accurately maintain the productivity of the crop and reduce economic losses. Foliar diseases are a special concern in peach because they can cause yield as well as quality if not timely detected. Artificial intelligence, machine learning and deep learning are some of the advanced technologies that are gaining great importance in today's agriculture, especially with the image analysis applications. In this study, a deep learning method for automated detection and classification of three peach leaf diseases and healthy class was presented based on image data. The dataset were taken at different phenological and disease stages under temperate conditions in Kashmir with four classes Healthy, Leaf Curl, Shot hole and Rust. Three convolutional neural networks (CNNs) architectures were applied, VGG-16, ResNet 50 and Xception were trained using transfer learning and Inception-V4 was trained from scratch for a comparative study of the learning strategies. Data augmentation techniques were applied to improve generalization. Results show that all models were able to learn disease specific features well. The result of Inception-V4 was found to be highest with 97.50%, followed by ResNet-50 with 94.49%, VGG-16 with 92.04% and Xception with 75.95%. The results of transfer learning-based architectures were also good and competitive but the best results obtained from the Inception-V4 architecture reveal its capability in modelling complex visual patterns. The results highlight the potential of deep learning techniques for early detection of diseases in peach, supporting precision agriculture and better disease management.
Accurate and non-destructive assessment of drought stress is important for improving lettuce production and supporting timely crop management. This study presents a detection-guided deep learning framework for plant-level drought-stress assessment in hydroponically grown lettuce using bird’s-eye-view RGB images. The study further investigates whether canopy segmentation can improve classification performance by reducing irrelevant background information. The framework was evaluated using 2190 images collected across three independent cultivation cycles in which drought stress was induced by isolating the plant root zones from the nutrient solution. In the first stage, YOLO-based object detection was used to localize individual plants, with YOLO26m achieving the highest detection performance of 99.4% mAP@0.5. The detected regions were subsequently used as spatial prompts for zero-shot canopy segmentation using the Segment Anything Model (SAM), with SAM ViT-B achieving a mean IoU of 0.9864. Six convolutional, transformer-based, and hybrid classification architectures were then evaluated independently using YOLO-cropped and SAM-segmented plant images. Segmented inputs consistently improved classification performance, with MaxViT-S achieving the highest binary test accuracy of 96.3%. The framework further distinguished time-defined pre-stress, early-stress, and late-stress periods with an accuracy of 92.4%. Plant-level generalization was further assessed using six-fold leave-one-plant-out cross-validation, resulting in a mean test accuracy of 90.25 ± 1.78% on unseen plants. These findings demonstrate that RGB-based plant-level analysis can support non-destructive drought-stress assessment and that canopy segmentation improves classification by reducing background influence.
Background Global soybean production is constrained by scarce arable land, and standardized evaluation tools remain lacking for natural mixed saline-alkali stress, the predominant abiotic stress under field conditions. Objective This study aimed to establish a comprehensive saline-alkali tolerance evaluation system for soybean germplasms via integrated multivariate statistical methods, and screen core and auxiliary indicators for efficient germplasm identification. Methods Seventy-one soybean germplasms were tested under 90 mmol/L mixed saline-alkali stress (NaCl:Na 2 SO 4 :NaHCO 3 :Na 2 CO 3 = 1:9:9:1, pH 8.2) simulating natural saline-alkali soil. We quantified the saline-alkali tolerance coefficients (SATC) of 13 morphological and physiological indicators, followed by coefficient of variation (CV), principal component analysis (PCA), subordinate function, cluster analysis and regression modeling. Results Significant inter-germplasm variations in saline-alkali tolerance were detected, and indicators with CV > 0.35 ( e.g ., root length (RL), root fresh weight (RFW)) were screened as primary indices. PCA extracted five principal components with 87.18% cumulative variance contribution, and the integrated analytical pipeline categorized germplasms into five tolerance grades: eight highly tolerant, 24 moderately tolerant, 13 generally tolerant, 16 sensitive and 10 highly sensitive accessions. A high-precision prediction model was constructed ( D = 0.290 X 1 - 0.026 X 2 + 0.438 X 3 + 0.402 X 4 + 0.180 X 5 + 0.153 X 6 + 0.813 X 7 - 1.123; R 2 = 0.998, where X 1 - X 7 represent the SATC of germination rate (GR), RL, RFW, total fresh weight (TFW), shoot dry weight (SDW), root dry weight (RDW), and total dry weight (TDW), respectively). A Chi-squared Automatic Interaction Detection (CHAID) decision tree model was further developed and validated using 10-fold cross-validation, yielding a cross-validation risk value of 0.003, which was comparable to the resubstitution risk value (0.002), indicating good generalization ability and low risk of overfitting. A novel five-dimensional overlapping analysis identified RFW as the core evaluation indicator, with RDW, TFW and R/S as key auxiliary indicators. Conclusion This study delivers a standardized, reproducible technical framework for large-scale screening of saline-alkali-tolerant soybean germplasms. It facilitates global saline-alkali land utilization, accelerates worldwide soybean stress-tolerance breeding, and provides a transferable paradigm for stress tolerance evaluation in other major crops.
Abstract Purpose Timely, accurate, and field-scale crop yield mapping is essential for precision crop management, yet most existing studies rely on late- or full-season data, limiting in-season decision-making. This study aims to develop a stage-aware earliest possible corn yield mapping framework that balances early data availability, predictive accuracy, and spatial fidelity by integrating Sentinel-2 imagery, vegetation indices (VIs), and accumulated growing degree days (AGDD). Methods Corn yield data were collected from a commercial farm over three growing seasons (2018–2020). The final modeling dataset included 51,794 Sentinel-2 10 m aggregated yield samples across three seasons: 2018 ( n = 16400), 2019 ( n = 18534), and 2020 ( n = 16860). Sentinel-2 raw spectral bands and derived VIs were organized for V4, V6, R1, R5, and R6 growth stages based on AGDD- and DAP-defined corn growth-stage windows, also confirmed visually on-ground, allowing observations from different planting and harvest dates across the three growing seasons to be aligned by crop developmental stage rather than calendar date. A stage-wise Pearson correlation and frequency-based selection identified informative and non-redundant VI subsets across crop development. Four machine learning models: Random Forest (RF), XGBoost (XGB), k-Nearest Neighbors (kNN), and a Neural Network (NNET; multi-layer perceptron) were trained using 13 input configurations, including individual growth stages and multi-stage combinations capturing phenological progression. Models were tuned via randomized search, trained on 2018–2019 data, and independently validated on the 2020 season. Yield predictions were mapped directly at 10 m Sentinel-2-pixel resolution without spatial interpolation to preserve fine-scale variability. Results Model performance was strongly influenced by phenological stage selection. Among single-stage inputs, R1 was the earliest stage where reliable yield mapping could be availed (RF: R² = 0.56, RMSE = 29.50%). While combining V6 with R1 stage inputs substantially improved predictive performance (RF: R² = 0.70, RMSE = 24.13%) for the yield mapping at the R1 stage, where V6-stage signals provided complementary yield-related information. The full-season combination (V4 + V6 + R1 + R5 + R6) produced the highest accuracy (RF: R² = 0.72, RMSE = 23.28%) but would be less suitable for early in-season decision-making. Early- or late-stage-only inputs (V4, R5, R6) showed weaker and less stable cross-year performance. Among algorithms, RF consistently generalized best to the independent 2020 dataset and is recommended for operational use. XGB showed strong training performance but reduced cross-year stability, kNN yielded moderate accuracy, and NNET achieved accuracy comparable to RF while closely reproducing observed spatial patterns such as center-pivot effects and edge gradients. Direct 10 m mapping preserved yield heterogeneity and avoided smoothing artifacts common in interpolation-based approaches. Conclusion Stage-aware feature selection and phenology-informed input combinations are critical for balancing yield prediction timeliness and accuracy. For operational in-season yield mapping, the RF model using the V6 + R1 stage combination provides a practical and reliable solution, enabling early, accurate, and spatially detailed yield estimates with robust cross-year performance. This study presents a deployable framework for integrating satellite time series and weather data into conventional (non-sequential) machine learning models to support proactive, within-season decision-making in precision agriculture.
Conventional breeding for ideotypes in target environments remains challenging due to genotype-by-environment interactions and the genetic complexity of key agronomic traits. Traditional multi-environment field trials are costly and time-consuming, limiting rapid genetic gain. These challenges highlight the need for digital tools to support rice breeding. However, two major approaches, genomic prediction (GP) and gene-based crop models (GBCMs), have distinct advantages. In this study, a dataset derived from a natural rice population comprising 210 genotypes, genotyped with about 650,000 markers, rice dry matter, and yield across three environments, was used to develop two genomic prediction models, genomic best linear unbiased prediction (GBLUP) and a convolutional neural network (CNN), together with a gene-based crop modeling framework. The effectiveness of these models in predicting rice traits and assisting in breeding selection was subsequently evaluated. Prediction results indicated that biomass and yield could be effectively predicted by all models, with Normalized Root Mean Square Error (NRMSE) values ranging from 10.60% to 18.59% and 9.93% to 18.19%, respectively. In terms of predictive accuracy, parameter-based crop models achieved the highest predictive accuracy, although it was confined to theoretical simulations. This was followed by the GBCM and CNN, whereas the GBLUP exhibited the lowest performance. Furthermore, GGE biplot analysis revealed the predictions of the GBCM aligned more closely with field observations than those of the CNN, emphasizing the potential of GBCM as a practical surrogate for digital breeding. These results provide valuable insights into modeling genotype-by-environment interactions and support the development of data-informed breeding strategies for future rice improvement.
Accurate long-term monitoring of winter wheat phenology is important for crop growth assessment and irrigation management, but 30 m Landsat-based phenology retrieval remains challenging because of sparse observations, cloud contamination and sensor differences. This study developed and evaluated an integrated Landsat-based workflow for monitoring winter wheat phenology in the People’s Victory Canal (PVC) Irrigation Area of northern Henan and the Alar Irrigation Area of southern Xinjiang from 2000 to 2024. Winter wheat areas were mapped using temporally stacked NDVI/EVI features and a CART classifier. Vegetation-index trajectories were reconstructed using locally adjusted cubic-spline capping combined with Savitzky–Golay filtering, and green-up, jointing, heading and maturity were extracted using threshold- and derivative-based detection. The CART-based mapping achieved an overall accuracy of 89.51%, with higher accuracy in Alar (91.45%) than in PVC (84.35%). Compared with S-G-only, Whittaker and TIMESAT-like approaches, LACC + S-G reduced phenological-date errors, especially for green-up and maturity, with RMSE values within 3.1 d against agro-meteorological observations. Phenological stages generally occurred later in Alar than in PVC, and spatial autocorrelation confirmed significant clustering. Agro-meteorological analysis suggested stronger thermal associations in PVC and stronger moisture-related associations in Alar. These results provide practical 30 m phenological information for regional winter wheat monitoring and irrigation scheduling analysis.
Lodging is a major yield-limiting factor in soybean, but efficient large-scale phenotyping and genetic dissection of this complex trait remain challenging for breeding programs. To bridge this gap, this study developed an integrated, breeding-oriented framework that links UAV-based high-throughput phenotyping with candidate gene identification. Field experiments involving 741 diverse soybean genotypes were conducted over two years, with UAV remote sensing performed at key reproductive stages (from R5 to R7). We identified UAV-derived structural (relative plant height), textural (homogeneity, dissimilarity, correlation), and spectral (NDVI, EVI, NDRE) features as the most sensitive indices for retrieving lodging severity. The fusion of these complementary features, coupled with the XGBoost algorithm, achieved high classification accuracy (0.81–0.92) across genotypes, growth stages, and years. This reliable phenotyping pipeline enabled the precise selection of contrasting genotypes (lodging-resistant vs. lodging-prone) for transcriptomic analysis. Transcriptome sequencing revealed 13,447 differentially expressed genes, with significant enrichment in phenylpropanoid and starch–sucrose metabolic pathways. Moreover, the haplotype analysis within a natural population identified superior allelic variants of two candidate genes ( Glyma.19G249100 and Glyma.05G142200 ) significantly associated with soybean lodging resistance. This work can effectively bridge the gap between scalable field phenotyping and the discovery of functionally validated breeding targets, providing an efficient and translational framework to accelerate the development of lodging-resistant soybean varieties.
Canopy 3D architecture is a critical determinant of light interception, photosynthetic efficiency, and final yield in cotton, yet its rapid and accurate characterisation remains challenging in field conditions. To achieve efficient, non-destructive, and quantitative monitoring of the canopy structure of field-grown cotton, this study proposes a 3D structure-based technology stack for high-efficiency, low-cost, and high-precision phenotyping extraction. This stack directly addresses the technical bottlenecks of traditional 3D data acquisition, namely high cost, long processing time, and low operational efficiency, which have hindered large-scale application. We developed a pipeline that integrates a fast reconstruction algorithm with a scale-recovery mechanism using ground control points (GCPs), enabling the generation of true-scale 3D point clouds from UAV aerial images in a cost- and time-effective manner. Using only 141 UAV images and with a reconstruction time of approximately 20 min, we efficiently reconstructed high-quality, scale-accurate point clouds of two 5.5 m × 5.5 m cotton plots, significantly outperforming SfM-MVS and Instant-NGP in terms of both reconstruction efficiency and point cloud completeness. This method, whose current validation is confined to a single season, one growth stage, and two experimental plots, not only achieves a breakthrough by using fewer input images with high efficiency, but also ensures point cloud accuracy and completeness, showing strong potential for rapid field monitoring and real-time management. Based on the high-quality reconstructed point clouds, we further quantitatively evaluated canopy characteristics at harvest, analyzing the coefficient of variation of canopy height, porosity distribution, and canopy volume fraction. The core shortcomings and optimization strategies for the existing canopy structure were identified, providing scientific data support and practical technical references for precision cultivation management and mechanization-compatible planting in cotton.
Accurate estimation of the cotton seedling Leaf Area Index (LAI) is essential for yield prediction and precision crop management. Traditional manual methods are limited by low throughput, while two-dimensional remote sensing techniques often struggle with sparse canopy cover and soil background interference. Although three-dimensional LiDAR presents a promising alternative, existing deep learning approaches typically depend on costly fully supervised point-wise annotations. To address this challenge, this study proposes an end-to-end framework that integrates UAV-based LiDAR, weakly supervised segmentation, and physical parameter inversion. A weakly supervised network, termed CogNet, was developed—incorporating self-distillation and structure-aware label propagation—to achieve precise segmentation of cotton plants using only 10% sparse annotations. Following instance segmentation via Density-Based Spatial Clustering of Applications with Noise (DBSCAN), individual plant phenotypic traits were extracted. A nonlinear Extreme Gradient Boosting (XGBoost) model was then constructed to invert LAI by leveraging allometric relationships between 3D structural parameters and leaf area. Experimental results showed that CogNet achieved an Intersection over Union (IoU) of 85.34%, effectively mitigating overfitting to label noise and achieving performance competitive with the fully supervised RandLA-Net (82.13%). Notably, under the specific conditions of this cotton seedling dataset characterized by strong geometric priors, the weakly supervised model demonstrated enhanced robustness against annotation inconsistencies. The framework attained a plant detection rate of 96.2%, and the XGBoost model delivered high estimation accuracy (R² = 0.879, RMSE = 0.138). This study demonstrates that weakly supervised learning can substantially reduce annotation costs while maintaining model performance, providing an efficient and cost-effective solution for field-scale crop phenotyping.
Sun-induced chlorophyll fluorescence (SIF) is an effective proxy for vegetation photosynthesis, but tower-based retrieval suffers from atmospheric path interference under humid and variable conditions. We present a DOAS-based SIF retrieval algorithm that operates in Fraunhofer lines (680–686 nm, 745–758 nm) and water vapour-sensitive bands (717–727 nm). It constructs an adaptive reference spectrum from SCOPE simulations and PCA and incorporates H2O absorption cross-sections into the fitting process for active atmospheric correction. The algorithm is implemented in a dedicated tower-based system integrating a 1° scanning gimbal with a high-resolution spectrometer. Validation with simulated and field data demonstrates the following: (1) the algorithm retrieves SIF with high fidelity (correlation coefficients >0.9 across all windows); (2) it exhibits lower water-vapour sensitivity and greater cloudy-sky stability than FLD, 3FLD, and SFM, achieving the lowest coefficient of variation (CV = 0.356); (3) over a complete wheat–rice rotation, the retrieved SIF tracks crop growth and phenological stages. This work provides a reliable solution for automated, high-precision tower-based SIF observation under complex atmospheric conditions.
Manisha Das Chaity · Ramesh Bhatta · Byron Eng · Jan van Aardt
Field / plotPhotogrammetry / SfM / MVSLiDAR / point cloudMultispectral / hyperspectralWhole plant / canopy / plot / field2D/3D reconstruction
The Greater Cape Floristic Region (GCFR) in South Africa is a fire-prone biodiversity hotspot where high species richness, structural complexity, and small plant sizes (0.0001–4 m2) pose substantial challenges for remote sensing-based biodiversity assessment. Spectral similarity among species and the mismatch between plant size and sensor pixel dimensions limit the capacity of current and forthcoming spaceborne systems to resolve individual species and accurately detect plot-level diversity changes. We therefore developed a physics-based simulation framework that couples fynbos trait measurements with radiative transfer modeling in the DIRSIG (Digital Imaging and Remote Sensing Image Generation) environment towards quantifying information loss across spectral and spatial scales and to define theoretical limits for biodiversity monitoring. We constructed a three-dimensional virtual scene of post-fire fynbos communities in Grootbos Private Nature Reserve, integrating high-resolution imagery, terrestrial laser scanning (TLS), and structure-from-motion (SfM)-derived point clouds. Field measurements of mean diameter and percent cover were used to scale vegetation models and constrain species abundance. We distributed plant instances using a blue noise sampling algorithm, guided by density maps derived from unmanned aerial system (UAS) imagery. Species-specific optical properties were parameterized using field-measured reflectance data and the PROSPECT radiative transfer model, while terrain structure was derived from SfM-based digital terrain models. The integrated scene was used to simulate multispectral (DJI Mavic 3 MSI), hyperspectral (AVIRIS-NG), and light detection and ranging (LiDAR) observations. Agreement between simulated outputs were evaluated against corresponding field-acquired datasets using spectral signatures and vegetation indices. This framework enables systematic assessment of sensor specification effects on spectral biodiversity metrics and provides a pathway for evaluating theoretical limits of species discrimination across airborne and satellite platforms.
Unmanned aerial vehicle (UAV)-based multispectral imaging enables flexible, non-destructive crop monitoring. Although UAV imagery offers much higher spatial resolution than satellite platforms, its effective spatial detail at typical operational flight altitudes can still be insufficient for plant-level interpretation and fine canopy structure, which can reduce vegetation-index reliability. Most super-resolution (SR) research targets RGB or satellite imagery and emphasizes perceptual or pixel-wise quality, leaving the spectral fidelity of reconstructed UAV multispectral imagery under-examined. This study benchmarked an SR evaluation framework for UAV-based five-band crop imagery (Blue, Green, Red, Red-edge, and near-infrared) using the open-source AI Hub cabbage dataset, with low-resolution inputs generated by controlled downsampling at ×2, ×3, and ×4. Nine methods (bicubic, SRCNN, EDSR, RCAN, SwinIR-based, ESRGAN-based, HAT-based, DAT-based, and DRCT-based SR) were compared under joint five-channel and band-wise reconstruction on 2170 test scenes using image-quality, spectral-angle, vegetation-index (NDVI, GNDVI, NDRE), band-wise, and efficiency metrics. EDSR and RCAN gave the most balanced performance. At ×4, band-wise reconstruction was strongest for per-band spatial fidelity, where EDSR reduced RMSE by 12.6%, and RCAN lowered near-infrared RMSE by about 21% relative to bicubic, whereas joint reconstruction with its spectral-angle and vegetation-index losses best preserved spectral relationships (spectral angle and vegetation-index errors). Learning-based gains were clearest at ×4. The recently proposed HAT-based, DAT-based, and DRCT-based attention models achieved the strongest pixel-wise RMSE and PSNR but did not surpass EDSR or RCAN on spectral angle or vegetation-index preservation under the equalized training budget. These results indicate that UAV multispectral SR should be assessed by spatial fidelity together with spectral consistency and agricultural index reliability.
Abstract In real field scenarios in agriculture, automatic segmentation of plant diseases is an important technique for precision farming. However, it remains exceptionally challenging due to blurred lesions, complex morphological structures, irregular backgrounds, and severe class imbalance. While traditional convo lutional networks struggle to capture long-range semantic context and standard vision transformers fail to preserve sharp localized boundaries, this paper proposes an efficient, attention-gated hybrid framework optimized for field deployment. Our architecture leverages a hierarchical Mix Transformer (MiT-B2) encoder stream integrated with an Atrous Spatial Pyramid Pooling (ASPP) scale-space context bridge and a custom Cross-Scale Multimodal Attention Gate (CMAG) to isolate discriminative disease markers selectively. Evaluated on the highly challenging and unbalanced PlantSeg dataset, our framework achieves competitive mean Intersection over Union (mIoU) of 66.57% and an F1-score of 79.93%, while maintaining a highly compact parameter footprint of only 30.37 M. Experimental evaluations demonstrate that the proposed system establishes a new performance milestone, outperforming current competitive architectures and proving highly viable for resource-constrained edge devices. To further enhance out-of-distribution stability, we outline future directions to extend our top-performing candidate variants into a Level 1 meta-stacking ensemble optimized via few-shot learning and partial backbone fine-tuning.
Functional-structural plant models simulate plant responses to environmental conditions, but their development and evaluation are often limited by the lack of datasets combining detailed architectural and physiological measurements. Here, we present a comprehensive dataset acquired from four oil palm plants ( Elaeis guinnensis) grown under controlled and contrasting climate scenarios. The dataset includes (i) three-dimensional reconstructions of plant architecture derived from terrestrial lidar point clouds, (ii) leaf-level gas exchange measurements used to parameterize photosynthesis and stomatal conductance models, and (iii) continuous plant-scale measurements of CO 2 and H 2 O fluxes obtained in a microcosm under precisely monitored and manipulated environmental conditions (light, temperature, humidity, and CO 2 concentration) across height climate scenarios. By combining detailed structural data with physiological measurements at both leaf and whole-plant scales, this database has been designed to build and evaluate digital twins (or shadows) of plants functioning under controlled conditions. It provides a valuable resource for calibrating biophysical models (light interception and photosynthesis), benchmarking model predictions across scales, and investigating the consistency between leaf-level parameterization and plant-level fluxes. All data and processing workflows are openly available, facilitating reuse for model development, evaluation, and intercomparison in plant and crop modelling communities.
Plant diseases remain a threat to global agricultural productivity, food security and livelihoods, especially in developing countries where the availability of experts in agriculture is still limited. The recent progress in AI, particularly deep learning and computer vision, has ushered in new possibilities for automated plant disease diagnosis, especially for plant image-based systems. This paper provides a systematic review of the deep learning methods employed for plant disease diagnosis, highlighting CNN-based methods, the application of transfer learning, explainable AI (XAI) methods and deployment issues. In the framework of PRISMA 2020, the relevant peer reviewed literature from 2016 to 2025 was systematically identified, screened and analysed on the most important academic databases. The review compared some of the most popular architectures such as GoogLeNet, DenseNet-121, MobileNetV2, EfficientNet, Attention-CNNs and Vision Transformers. Results showed very high classification accuracy in controlled lab conditions with DenseNet-121 achieving ~99.75% accuracy with good computational efficiency. But it also revealed a big gap between the lab and the field, mainly due to environmental variations, domain shifts, and dependence on datasets. Some innovative and emerging technologies like explainable AI, hyperspectral imaging, few-shot learning, and lightweight mobile architectures showed promise of enhancing the interpretability, early detection of disease, and the use of smart phones in low-resource agricultural settings. In conclusion, the study suggests that in order to be implementable in the field, future intelligent agricultural diagnosis systems must be able to balance predictive accuracy, explainability, computational efficiency and field adaptability. The results enrich the existing knowledge on precision agriculture and serve as useful information for researchers, agricultural technologists, and policymakers working on the creation of AI-based systems for crop protection.
Plant diseases lead to substantial yield losses and pose a persistent threat to global food security, creating an urgent demand for high-throughput, accurate, scalable, and non-destructive disease-monitoring approaches. Remote sensing has emerged as a powerful tool, yet progress in plant disease detection remains fragmented across various disciplines, tasks, sensing methods, and data modalities. This review introduces a hex-view perspective to synthesise remote-sensing-based plant disease detection within a cohesive conceptual framework. Instead of treating sensing technologies, algorithms, and datasets independently, the hex-view incorporates six interconnected dimensions that jointly capture how biological processes, the measurement scale, and data characteristics constrain disease detectability, including when detection is possible and how reliably it can be achieved. The hex-view framework comprises six interconnected dimensions and forms an integrated framework called BTSCAD: (1) Biology (B): plant–pathogen interactions constituting the biological foundation of disease development and expression. (2) Task (T): the diverse disease-detection tasks and their corresponding research objectives. (3) Sensor (S): the sensing modalities that define the data acquisition type and richness of captured information. (4) Condition (C): the environmental conditions, sensing platforms, and spatial scales that shape disease observations and bridge controlled experiments and real-world deployment across leaf, canopy, plot, and regional scales. (5) Algorithm (A): the classical and state-of-the-art data-analysis algorithms used to extract disease-related information from sensor data. (6) Dataset (D): the data sources that underpin model development, evaluation, and generalisability. The hex-view perspective provides a clear framework for interpreting previous research and identifying future research directions. This review lays a structured foundation for developing robust, interpretable, and transferable disease-detection systems, supporting advancements in precision agriculture, high-throughput phenotyping, and sustainable crop production.
Localized soil-moisture deficits, that is, irregular sub-field patches where crops experience water stress well before visible wilting, are a leading cause of yield variability in row-crop agriculture. These zones are difficult to detect at the spatial resolution and revisit frequency required for timely irrigation response. This paper presents a reinforcement-learningguided autonomous quadrotor unmanned aerial vehicle (UAV) platform that fuses onboard Visual Simultaneous Localization and Mapping (Visual SLAM) with a pushbroom hyperspectral imaging payload to construct georeferenced, canopy-registered maps of a Crop Water-Stress Index (CWSI) in near real time. Rather than flying a fixed lawnmower survey, the platform is guided by an adaptive-sampling policy trained with Proximal Policy Optimization (PPO) that reallocates flight time and sensor dwell toward regions of emerging water stress as evidence accumulates mid-flight. We present the complete engineering pipeline: airframe and sensor design, a keyframe-based Visual SLAM front and back end that provides centimeter-scale geolocation without continuous reliance on Real-Time Kinematic (RTK) GNSS lock, a hyperspectral preprocessing and spectralindex chain (NDVI, NDRE, NDWI/NDMI) used to derive CWSI through a learned regression, the partially observable Markov Decision Process (POMDP) formulation and reward shaping used to train the sampling policy, and the fused system architecture tying these subsystems together. In simulated field trials over a 0.8-hectare test plot, the reinforcement-learning-guided policy achieved a 92% water-stress-zone detection rate versus 61% for a fixed-grid baseline, while reducing mission flight time by approximately 32%. We further report an ablation study isolating the contribution of SLAM-derived canopy structure to CWSI accuracy, a sensitivity analysis across field complexity, and a full error budget for the fused pipeline. We close with a discussion of validation limitations, broader scientific and agricultural impact, and a roadmap toward multi-UAV fleet deployment for whole-farm monitoring
Abstract Visual assessments of growing forest nursery plants are time-consuming and often result in a lack of information at a physiological level. There exists a need for health screening in nurseries, that is fast and efficient, to improve overall health monitoring and nursery productivity. Rapid handheld sensors such as rapid thermal devices, leaf porometers and moisture meters, can provide regular information at a physiological level, that can improve the understanding of the impact of stress on young plant cuttings and their decline in health over time. This paper evaluates the utility and reliability of contemporary sensor technologies, to operationally monitor stress phases in juvenile forest plant cuttings during progressive moisture (dry-down) conditions. Furthermore, to assess whether thermal sensors could be used as an indicator, in conjunction with other variables such as soil water content or stomatal conductance, is needed operationally for fast screening during limited planting windows. Near Infra-Red Analysis (NiRA) data was collected to understand detailed plant functions at a finer reflectance level. A relationship was found where the increase in thermal signals reflects a depletion of water content, resulting in an eventual decline in stomatal conductance and, ultimately, plant mortality. Several algorithms were used in a preliminary test, using RapidMiner software, to discriminate between the four phases of plant health decline using physiological variables and NiRA data. Both Gradient Boosting Trees (GBT) and Deep Learning (DL) showed the best performances, achieving favourable accuracies of 96.8% and 91.2% without NiRA data, 84.6% and 88.2% with NiRA data, with shorter training times. Using thermal technology weighted amongst the highest of the best performing variables using GBT, the utility and accuracy showed good discrimination between the stages of plant decline and is encouraged for future research in this field.
Mateo Pastrana · C. Velilla · Nelson Mattie · Alfonso Gomez · Sergio Molina
Field / plotLiDAR / point cloudWhole plant / canopy / plot / field2D/3D reconstructionYield / biomass estimationBiomass / plant weight
Reliable aboveground biomass (AGB) estimates for woody crops are essential for carbon accounting and for Measurement, Reporting and Verification (MRV) frameworks. However, it remains unclear how LiDAR modality and sampling geometry influence plot-scale and tree-scale AGB predictions in intensively managed Mediterranean orchards. In this study, we benchmarked four LiDAR modalities, namely open national airborne laser scanning from the Spanish National Aerial Orthophotography Plan (PNOA/ALS), a dedicated Riegl airborne laser scanner (ALS), unmanned laser scanning (ULS) and mobile laser scanning (MLS), across three woody-crop sites in Córdoba (southern Spain): IFAPA, Doña María, and Villaseca. Plot-level LiDAR metrics (mean height, 95th height percentile, maximum height, and canopy-cover proxies) were extracted from normalized point clouds and related to field AGB using Random Forest and XGBoost regression models, together with an ensemble predictor, under an 80/20 train–test split. In parallel, TreeQSM-based Quantitative Structure Models (QSMs) were evaluated as an independent tree-level three-dimensional reconstruction approach. XGBoost achieved the lowest errors at IFAPA (RMSE = 0.400 Mg ha−1; R2 = 0.994) and Villaseca (RMSE = 0.872 Mg ha−1; R2 = 0.995), whereas PNOA/ALS was competitive at Doña María (RMSE = 0.725 Mg ha−1; R2 = 0.994). TreeQSM closely matched the field inventory at the low-biomass IFAPA site but tended to overestimate biomass at Doña María and Villaseca, and only 28% of scanned trees yielded usable reconstructions. The results support the use of cross-platform LiDAR for orchard AGB and carbon mapping and identify the conditions under which open national LiDAR can enable scalable MRV of Mediterranean woody crops.
Soil salinity is a major constraint on rice ( Oryza sativa L.) production, particularly during the yield-determining reproductive stage. Utilizing a high-throughput RGB platform, we non-destructively phenotyped a diverse panel of 294 rice accessions over two consecutive years. By extracting 60 dynamic image-based traits (i-traits) reflecting canopy architecture and stay-green capacity, and four seed-setting rate-related traits, our genome-wide association study (GWAS) identified 95 significant loci, 35.8% of which precisely co-localized with previously reported QTLs. We further prioritized OsSLT1 ( LOC_Os01g05790 ) as a candidate gene at a reproducible suggestive locus associated with leaf-rolling-related image traits. Transgenic evaluations confirmed that it acts as a positive regulator of salt tolerance at the seedling stage. OsSLT1 was mainly detected in the nucleus, and no significant changes in Na + or K + accumulation were observed in flag leaves under the tested salt-stress condition, suggesting that OsSLT1 may regulate salt tolerance through mechanisms beyond classical shoot ion accumulation. Natural variations in the OsSLT1 promoter were associated with transcriptional divergence. The Hap2 promoter haplotype showed significantly higher stress-induced transcriptional activity. Hap2 was rare in modern indica accessions, suggesting that it may represent a potentially useful genetic resource for future salt-tolerance improvement. Supplementary information The online version contains supplementary material available at 10.1007/s11032-026-01704-2.
Plant phenotyping is essential for modern crop breeding, yet traditional static image analysis fails to capture the nonlinear dynamics of plant growth. Existing time-series forecasting models exhibit notable limitations when processing multimodal data: global pooling operations may compress local 2D spatial topology of plants, and shallow feature concatenation may be insufficient for effective cross-modal semantic alignment. Moreover, current methods typically regress absolute morphological states, which may contribute to temporal lag during nonlinear growth spurts. In this paper, we propose ST-CrossGro-Former, a cross-modal residual forecasting framework for plant dynamic growth. The network removes the final global pooling and classification layers to preserve spatial topology and incorporates a scalar-guided cross-modal attention module based on the standard query-key-value formulation. This module utilizes 1D morphological features as queries to dynamically weight local visual regions, promoting multimodal feature alignment. Concurrently, a residual incremental forecasting strategy is introduced to predict short-term growth increments rather than absolute states, aiming to improve tracking sensitivity to sudden growth events. Evaluations on the UNL-CPPD maize dataset and supplementary validation on the FIP1 wheat field dataset show that the proposed model achieves competitive single-step forecasting accuracy and favorable temporal trajectory alignment compared with adapted spatiotemporal attention, graph-based, and physics-informed baselines under the evaluated settings. In particular, the FIP1 results suggest that ST-CrossGro-Former can maintain favorable height trajectory alignment under a field-acquired wheat setting, indicating its potential for helping mitigate temporal misalignment in dynamic growth forecasting.
Reproduction assets foundThe paper evaluates its ST-CrossGro-Former model on two public plant phenotyping datasets: the UNL-CPPD maize dataset (explicitly stated as publicly available with a repository URL) and the FIP1 wheat field dataset (public dataset from ETH Zürich, with its GigaScience dataset publication DOI). No author analysis code, Dataset · publicThe UNL-CPPD dataset used in this research was acquired from the UNL Plant Phenotyping Datasets repository, accessible at https://plantvision.unl.edu/datasets.Open asset ↗UNL Plant Phenotyping Datasets · UNL-CPPDlines:266-273Plant phenotyping relevance matchCrossref · checked 14 Sept 2026
In Nigeria, maize is the most widely cultivated grain, largely supporting food security for about half of the population. Research has indicated that there is annual production is declaiming to approximately 50% in maize, this is due to crop diseases and pest damage. This research presents a deep learning surveillance system for crop disease prediction and pesticides recommendations based on a DenseNet-121 architecture for continuous video streams. The research was evaluated on a curated field dataset collected from three states in Nigeria; Adamawa, Borno, and Taraba State. The system achieved a mean accuracy of 98.2% and a mean F1-score of 0.982. The results reflect a strong discriminative capacity across the diverse textural maize diseases.
Manual inspection of grain plant leaves for defects is subjective and labor-intensive. Few studies have compared deep learning methods on a combined multi-crop dataset. The study collected locally 5,640 leaf images from rice, maize, and guinea corn farms in Nigeria and grouped them into six classes representing defective and healthy leaves for each crop. Three models were trained: YOLOv8 for end-to-end detection and classification, EfficientNetB0 for standalone image classification, and a hybrid that used YOLOv8 for leaf detection followed by EfficientNetB0 for patch classification. The hybrid achieved 99.85% accuracy on the test set, slightly above EfficientNetB0 (99.82%) and YOLOv8 (mAP 0.995). The hybrid also supplies bounding box locations, helping farmers identify exactly where damage appears. This system offers a reliable, field-deployable tool for monitoring grain crop health.
Red crown rot of soybean (RCR), caused by Calonectria ilicicola, is an emerging soilborne disease whose quantification is challenging due to its complex symptom development across root and foliage levels. This study developed and evaluated a multi-scale framework to improve the assessment of RCR severity from controlled environments to field conditions using root imaging and standardized visual scales. Under controlled conditions, a standard area diagram (SAD) for root necrosis was developed and validated, and SAD-assisted evaluations significantly improved accuracy, precision, and inter-rater agreement compared with unaided assessments. In field conditions, a diagrammatic symptom scale (DSS) was developed using consensus-rated images from experts and showed high reliability, repeatability, and reproducibility across 18 raters, with strong intra- and inter-rater agreement. This study developed and evaluated complementary methods to improve the assessment of RCR severity from controlled environments to field conditions using root imaging and standardized visual scales.
Quantitative flowering phenotypes are needed to support breeding and harvest management in Hypericum perforatum L. (St. John’s wort), but manual flower assessment is slow and difficult to standardize under field conditions. UAV-Hyp is a multi-temporal UAV RGB dataset containing 12,653 high-resolution images acquired at 26 measurement dates across the complete flowering period of 15 H. perforatum accessions. The images represent variable illumination, soil moisture, weed pressure, and developmental stages. The dataset provides 59,163 plant bounding boxes and 107,054 flower bounding boxes. As an application example, cascaded YOLOv8 plant and flower detectors achieved mAP@0.50:0.95 values of 0.977 and 0.950, respectively. UAV-Hyp supports scalable flower quantification and the development of time-series phenotyping methods for genotype comparison and quality-oriented medicinal-plant breeding.
Antonio Samudio Oggero · Daisy Leticia Ramírez Monzón · Héctor D. Nakayama · Luis Felipe Medeiro Alves · Valter Arthur · Oscar Vega Alvarenga · G. Resquín-Romero · Wilson Romero Vergara · Juan D. Avalos Añazco
Calibrating the mutagenic dose is the first practical step of any radiation mutation-breeding programme, and it is usually summarised by the median lethal dose (LD50) or the median growth-reduction dose (GR50). We asked whether an accessible, image-based phenotyping pipeline can quantify the early radiation response of cowpea (Vigna unguiculata L. Walp.) seedlings finely enough to estimate GR50 and to rank organ- and pigment-level sensitivities. Seeds of the traditional Paraguayan landrace kumandá pyta’i were exposed to Cobalt-60 gamma rays at 0, 100, 200, 300, 400, 500, 600, and 700 Gy, grown in a greenhouse, and photographed at the early seedling stage. A single calibrated photograph (5.1 px mm−1) of 83 seedlings was segmented in Fiji/ImageJ 1.54p and analysed with Python to extract morphometric traits (total, root, and shoot length, root:shoot ratio, tortuosity, and a two-dimensional biomass proxy) and colorimetric traits (CIE L*a*b*, a normalised greenness index, and colour-class pixel fractions). Because the data departed from normality, dose effects were tested with Kruskal–Wallis, Spearman rank correlation, and Dunn post hoc tests, and GR50 was estimated by regression of each trait expressed as a percentage of the control. Total length, shoot length, and the biomass proxy declined significantly with dose (Spearman ρ = −0.40, −0.51, and −0.47; all p < 0.001), preceded by a low-dose stimulation at 100 Gy. Estimated GR50 values were ≈390 Gy for shoot length, ≈510 Gy for total length, and ≈550 Gy for the biomass proxy, within the range reported for other cowpea genotypes. Shoot elongation was more radiosensitive than root elongation, so the root:shoot ratio did not decline; tortuosity showed no dose response. Among pigment traits, the loss of greenness was the most robust signal (a* increased, ρ = +0.62, p = 5 × 10−10; green pixel fraction fell from 0.32 to near zero by 500 Gy). These results show that single-photograph phenotyping resolves a coherent, statistically supported dose response and yields a GR50 estimate usable for dose calibration. For kumandá pyta’i, doses of roughly 300–400 Gy (below GR50) are the most defensible starting window for mutation induction. The framework is reproducible and low-cost, but it is based on one greenhouse experiment and a single genotype, and should be validated across independent trials and cultivars.
Drought stress severely limits foxtail millet yield and quality, yet current drought-resistance indices are exclusively yield-oriented and ignore grain-filling quality. Our two-year (2024–2025) experiments with 24–48 varieties revealed that yield and blighted grain rate (BGR) are partially decoupled (e.g., Zhangzagu 18: yield 2307 kg/ha, BGR 0.444; Zhonggu 19: yield 1622 kg/ha, BGR 0.280). We therefore constructed the Yield–Quality Synergy Index (YQSI = DYI − BGR), which penalizes varieties with poor grain filling. The YQSI tied for first place with DYI in comprehensive screening performance and achieved the highest inter-annual stability (Spearman ρ = 0.823, Jaccard = 0.438, composite score = 1.261). Sensitivity analysis confirmed robustness of the equal-weight formula across a 4-fold range of quality-penalty weights. Six strongly drought-resistant germplasms with balanced yield and quality were identified. Using UAV multimodal data (RGB, multispectral, and thermal infrared) acquired during grain filling, a Random Forest model predicted a YQSI with overall R2 = 0.819 and an F1 score of 0.933 for variety screening. Feature-importance analysis highlighted NDVI, WDRVI, and red-edge texture as key predictors. This study provides a quality-constrained drought-resistance evaluation framework and demonstrates the potential of UAV-based high-throughput phenotyping for foxtail millet breeding.
This study investigates the automatic classification of crop phenological stages from low-altitude UAV RGB imagery. The dataset included 11,489 images of five crops: sunflower, rapeseed, soybean, wheat, and barley. The images were annotated using the Biologische Bundesanstalt, Bundessortenamt und Chemische Industrie (BBCH) scale, with labels corresponding to either single stages or stage ranges to reflect heterogeneous field conditions and transitional crop states. A pretrained ResNet18 model was adapted to the task using transfer learning. Training was conducted in two stages: first, the classification head was optimized while the backbone remained frozen; second, the entire network was fine-tuned. The model achieved strong internal test accuracy across all crops, with 100% test accuracy for rapeseed and barley, more than 99% for the remaining crops, and a mean accuracy of 99.73% under the studied survey conditions. The results also compare favorably with previously reported studies on UAV-based phenological classification. Overall, the findings support the potential of low-altitude UAV imagery and deep learning for localized phenological assessment of selected field zones in precision agriculture, while broader deployment requires validation across independent fields, seasons, regions, and survey conditions.
Matuszynska, A. · Sansa, O. · Adekoya, F. J. · Akinyemi, O. O. · Anokye, E. · Bashir, O. B. · Boyny, Z. Z. F. · Chukwuka, M. K. · Corvest, E. · Dada, A. O. · DellAcqua, M. · Ehemba, G. L. · Finkbeiner, A. J. · Hamabwe, S. · Hodehou, D. A. T. · Kacheyo, O. · Kamfwa, K. · Mhango, K. J. · Abdullahi, W. M. · Munduwe, G. · Ntukidem, S. · Obisesan, O. K. · Odesina, I. S. · Ogechi, N.-U. · Olaoye, O. D. · Olayinka, M. M. · Osei-Bonsu, I. · Rilwan, K. O. · Stival, L. · Tehar, Z. · Tende, R. M. · To, J. · Ugochukwu, U. K. · Unger, A. · van Aalst, M. · Vrbic, D. · Zhang, C. · Theeuwen, T. P. J. M. · Kramer, D. M. · Kromdijk, J.
Photosynthesis is among the most consequential yet genetically complex traits in crop plants, and translating its natural variation into actionable genomic targets remains a central challenge for breeding climate-resilient varieties. To start addressing this, researchers are generating increasingly large, multi-environment field photosynthesis datasets. Yet, these data have been structurally under-analysed since their inception. Here we report the outcomes of the first dedicated hackathon focused on computational mining of such field data held in Accra, Ghana, in March 2026. Bringing together data scientists, plant physiologists, geneticists, and breeders from Europe and Africa, these interdisciplinary teams used photosynthetic data collected with hand-held fluorometers to genome-wide marker data across four crop species: cowpea (Vigna unguiculata), barley (Hordeum vulgare), common bean (Phaseolus vulgaris), and potato (Solanum tuberosum). Despite using different species and methods, independent teams identified the same three key findings. First, mechanism-informed feature engineering and dynamic modelling recover genetic signals that are not detected or discarded in standard analysis pipelines, resulting in traits with improved heritability and meaningful associations with yield. Secondly, machine learning methods proved effective at uncovering genetic associations, with temporally resolved features substantially outperforming single time-point measurements. Third, raw chlorophyll fluorescence and absorbance traces consistently contained more information and predictive power than the extracted parameters currently used. A defining feature of this event was having experimentalists and data scientists working together, enabling AI approaches to be grounded in domain knowledge and biological mechanisms rather than relying on data alone.
The uniformity of maize plant spacing serves as a critical indicator for assessing sowing quality, seed vigor, and field seedling emergence stability. However, manual measurement is inefficient, and complex field conditions make automatic seedling detection and plant spacing measurement challenging. Aiming at the challenges of missed detection, insufficient accuracy for small targets, and large errors in plant spacing calculation under complex field conditions, this study constructs a high-quality dataset containing 693 maize seedling images and implements preprocessing enhancement for images degraded by haze or dust. An intelligent maize seedling detection and plant spacing measurement method based on improved YOLOv8 is proposed. The Global Attention Mechanism (GAM) is embedded into the backbone network to strengthen cross-dimension information interaction between channels and spaces, suppress background interference, and reduce the missed detection rate. The Bi-directional Feature Pyramid Network (BiFPN) is adopted to replace the original PAFPN for enhanced multi-scale feature fusion and deep semantic representation. A new 160 × 160 high-resolution small-object detection layer is added to significantly improve the detection performance of weak and small seedlings. Experimental results demonstrate that the improved model achieves a precision, recall, mAP50, and mAP50-95 of 89.4%, 90.3%, 94.4%, and 49.4%, respectively, which are 2.6, 0.5, 1.5, and 2.7 percentage points higher than those of the original YOLOv8 model. These results indicate that the proposed model improved maize seedling detection performance under complex field conditions. Automatic plant spacing calculation is realized based on detection outputs; the average plant spacing of the dataset is 30.42 cm, with a relative error of only 4.93% compared with the preset sowing spacing of 32 cm. The proposed method can efficiently accomplish field seedling identification, plant spacing quantification, and sowing quality evaluation, providing reliable technical support for precision maize sowing, seeder parameter optimization, and intelligent field management, which is of great significance for promoting the intelligent upgrading of grain crop production.
Abstract Purpose Long-term monitoring of crop biophysical and biochemical traits remains challenging in high-latitude regions due to short growing seasons, frequent cloud cover, and highly variable weather. In this context, unmanned aerial vehicles (UAVs) offer flexible, high-resolution observations, but their added value relative to low-cost proximal sensors and their effectiveness for radiative transfer model (RTM) inversion across diverse crop canopies remain insufficiently quantified. This study evaluated the potential of a two-band proximal spectral reflectance sensor (SRS) and a five-band multispectral UAV sensor for retrieving leaf area index (LAI), leaf chlorophyll content (LCC), and canopy chlorophyll content (CCC) using PROSAIL inversion across major crops in Northern Europe over two growing seasons (2023–2024). Methods and Results Two inversion approaches – look-up table (LUT) and artificial neural network (ANN) were applied to PROSAIL simulations. UAV–PROSAIL–ANN outperformed LUT-based inversion and SRS observations, achieving the highest accuracy for LAI (R 2 = 0.81–0.95; RMSE = 0.27–0.77 m 2 /m 2 ), followed by CCC (R 2 = 0.58–0.94; RMSE 2 ), while LCC remained less accurately estimated (R 2 = 0.26–0.78; RMSE 2 ). Across sensors and methods, retrieval accuracy decreased in the order of LAI, CCC, and LCC, reflecting the stronger spectral control of canopy structure compared to biochemical traits. Conclusions The UAV–PROSAIL–ANN framework effectively captured spatial and temporal variability in crop traits, producing canopy-scale maps consistent with field observations. These results demonstrate the robustness and scalability of hybrid PROSAIL–ANN inversion for high-latitude crop monitoring, while highlighting current limitations in biochemical trait retrieval using multispectral data.
Reproduction assets foundThe paper's Data availability statement explicitly deposits the authors' UAV image processing code (irradiance normalization, vignetting, exposure compensation, radiometric calibration) in a public GitHub repository under GPL v3.0; other data are available only upon request.Code · publicData availability Code to perform irradiance normalization, vignetting, exposure compensation, and radio-
metric calibration is available at https://github.com/fieldSITES/scripts/tree/main/UAV under GNU General
Public License v3.0. Other data will be made available upon request.Open asset ↗UAVpdf-page:34 lines:1-40Plant phenotyping relevance matchCrossref · checked 14 Sept 2026
Published14 Aug 2026Journal of Intelligent Decision Making and Information ScienceCited by 0 · OpenAlex ↗
This paper explores how deep learning methods can be used to monitor the health of crops and identify diseases, particularly for the apple crop. As the need for food security and sustainable farming methods increases, there is a strong demand for early detection of crop diseases. We have used a Convolutional Neural Network (CNN), based on the model of VGG16 architecture, since the model is known to be effective in image classification. The dataset contained 7771 training images for to enhance machines deep learning regarding plant diseases. Following that, validation image collection of 1747 images divided into four health conditions of the apple crops. In addition, the model was evaluated using 196 new images as a final test. To enhance the model capacity for recognizing diseases of real leaves rather than just memorize exact training pictures, data augmentation was used with ImageDataGenerator of TensorFlow. This means the training images were zoomed, rotated, and shifted to enable the machine detects more variations. Ten epochs of training were performed to measure the model accuracy. The results indicated that the model obtained significant improvement in training and validation accuracy from 56.43% to 78.12% and 92.94 to 96.93%, respectively. Most impressively, the final test dataset, which contained completely new images, scored an accuracy rate of 98%. The results indicate that in the architecture field the application of deep learning methodologies is effective, suggesting that automated detection of diseases by using sophisticated image analysis manages crop diseases identification efficiently. Combination of these methods successfully creates avenues for novel research to built real-time systems of crop disease monitoring, which help farmers increase their productions and farm their lands sustainably.
Accurate estimation of aboveground biomass (AGB) in dryland savanna woodlands is constrained by sparse field data, which has motivated widespread fusion of field plots with spaceborne LiDAR reference data from the Global Ecosystem Dynamics Investigation (GEDI). Here, we show that such fusion can substantially inflate apparent model accuracy when the two reference sources sample different spatial domains. Using 44 field plots from the Abu-Gadaf Natural Reserved Forest (AGNRF), Sudan, and 56 GEDI L4A footprints drawn from a 50 km buffer surrounding the reserve, we trained Random Forest (RF), Gradient Boosting (GB) and Classification and Regression Tree (CART) models on Sentinel-1, Sentinel-2, SRTM and Dynamic World predictors and evaluated them under 10-fold, 2 km block spatial cross-validation. The merged dataset yielded apparently moderate performance (RF: RMSE = 9.40 Mg ha−1, R2 = 0.33). However, GEDI-derived AGB was 2.1 times higher than field-measured AGB (18.71 vs. 8.89 Mg ha−1; Kolmogorov–Smirnov D = 0.53, p
High-throughput phenotyping is key in modern breeding for rapidly and cost-effectively evaluating salt-adaptive traits. However, few studies have combined spectral reflectance indices (SRIs) with deep learning to assess field-grown wheat under salt stress. In this study, we developed optimized 2D and 3D SRIs integrated with artificial neural network (ANN) to assess chlorophyll a (Chl a), chlorophyll b (Chl b), total chlorophyll (TChl), and grain yield (GY) in 32 recombinant inbred lines (RILs) and four cultivars under 150 mM NaCl field conditions. ANOVA revealed that genotype contributed the largest proportion of the treatment sum of squares across all traits (60–80%), followed by the genotype × year interaction (8–15%), whereas year contributed the smallest proportion (1–5%). Heatmap clustering of these traits clearly distinguished salt-tolerant from salt-sensitive genotypes. The study findings highlight using four key traits as screening criteria for salt tolerance in wheat. Our optimized 2D/3D spectral indices showed moderate to strong predictive power (R 2 = 0.25–0.75), outperforming earlier indices. Multi-season data improved accuracy by 15–25%, with best predictions for Chl a and TChl (R 2 = 0.34–0.75) versus Chl b and GY (R 2 = 0.25–0.64). Top models included ANN-3D-SRIs-8 for Chl a (R 2 = 0.735/0.644), ANN-3D-SRIs-3 for Chl b (R 2 = 0.611/0.549), ANN-2D-3D-SRIs-3 for TChl (R 2 = 0.713/0.619), and ANN-2D-SRIs-2 for GY (R 2 = 0.648/0.553). This framework combines optimized indices and machine learning for scalable, high-throughput phenotyping to advance precision breeding of salt-tolerant wheat.
This study evaluated the effects of input-channel composition and the incorporation of attention modules on the detection accuracy of visually suspected pine wilt disease (PWD)-affected trees using unmanned aerial vehicle (UAV) RGB orthomosaic imagery and the YOLO26-Large (YOLO26-L) deep learning model.Four datasets were constructed: Dataset A, consisting of three RGB channels; Dataset B, combining RGB with elevation and aspect (5 channels); Dataset C, combining RGB with three graylevel co-occurrence matrix (GLCM) texture features, namely Angular Second Mement (ASM), Entropy, and Homogeneity (6 channels); and Dataset D, combining RGB with the GLCM texture features and topographic information, namely elevation and aspect (8 channels).The Squeeze-and-Excitation (SE) block and Convolutional Block Attention Module (CBAM) were independently integrated into the YOLO26-L baseline model, and the detection performance of the 12 combinations was analyzed.In the accuracy assessment, the YOLO26-L baseline model trained with Dataset B achieved the highest mean Average Precision (mAP)@50 of 0.63 among the 12 combinations.However, the differences among the 12 combinations were marginal and did not indicate the superiority of a specific combination.Dataset C and Dataset D, which incorporated GLCM texture features, achieved accuracy levels similar to those of Dataset A, which used RGB alone, indicating that GLCM texture features did not substantially improve the detection of PWD-affected trees in 5-cm-resolution imagery.The application of attention modules also did not lead to a consistent improvement in accuracy, and the detection rates of all 12 combinations remained around 60% for objects smaller than 25 m².These results suggest that simply adding texture features or attention modules did not reliably improve detection accuracy.This study extends RGB-based UAV detection of PWD-affected trees by progressively integrating elevation, aspect, and GLCM texture features into four datasets and evaluating their interactions with SE and CBAM attention modules across a 12-combination experimental matrix.The results can inform the selection of input channels and model architectures in future forest disease detection studies.
Sustainable wheat farming is challenging. Real-time information on crop health, disease transmission, and anticipated yields is essential for farmers. However, they frequently use slow, expensive, or non-communicative tools. This project develops a workable solution. There is no need for massive server farms because the entire system operates on a single graphics card. It incorporates images of wheat fields, Indian farming notes, greenhouse records, harvest statistics, and NASA meteorological data. Consider them as various “eyes” for crop photo analysis, and we tried several lightweight computer vision models. ConvNeXt-Tiny was slower but could operate on older equipment with 75% accuracy; EfficientNetB0 recognised wheat heads with 92% accuracy; and AgroMark, a hybrid solution that merged photo analysis with agricultural metadata (soil type, rainfall, increased to 87%, etc. Combining picture analysis with attention mechanisms (CBAM) allowed us to anticipate the amount of wheat that a field will yield based on these photo insights, and the results showed that our predictions were accurate, with an R 2 score of 0.97. Additionally, we developed a versatile detector that simultaneously detects disease, stress, head count, and pests. It is adjusted to deal with training data that is unbalanced (some diseases are common, while others are rare). As we packed everything into a 16-GB graphics card, we spent real time determining which strategies smaller training sets, removing weak features, and adjusting loss functions, work. We encounter real-world obstacles along the road, such as photographs from different locations not always match, mislabeled photographs from different locations not always match, mislabeled diseases, and neglected rare pests. Our step-by-step instructions, charts, and code are available.
Reproduction assets foundThe paper builds its multimodal wheat phenotyping analysis on several explicitly cited public data assets: the Kaggle Wheat Plant Diseases image dataset (used for disease classification, Tables 2 and 9), the Global Wheat Head Detection dataset (used for head detection, Tables 1 and 6), FAOSTAT and India Open GovernmentDataset · publicAvailable online at: https://www.fao.org/faostat/ . FAOSTAT statistical database.Open asset ↗lines:1110-1162Plant phenotyping relevance matchEurope PMC · checked 5 Sept 2026
Abstract China's rice production and environmental sustainability are largely dependent on the cold black soil region in Northeast China, where precise water and nitrogen management is challenged by water scarcity and high carbon emissions. To overcome the limitations of conventional empirical management and improve the accuracy of evapotranspiration (ET) estimation in controlled-irrigation paddy fields, this study proposes a novel framework integrating unmanned aerial vehicle (UAV) multispectral and thermal infrared observations, the FAO-56 dual crop coefficient approach, and the NSGA-II multi-objective optimization model. To parameterize and validate this methodology, field data comprising four lower limit thresholds for controlled irrigation and four nitrogen fertilizer application rates were acquired from the Rice Research Site of Farm 856, Heilongjiang Province, China. This integrated approach was used to systematically evaluate rice growth, water consumption, resource use efficiency, and greenhouse gas emissions under different water-nitrogen treatments. Based on these evaluations, an irrigation optimization scheme was developed using daily crop evapotranspiration (ETc). The results indicated that water, nitrogen, and their interaction significantly affected rice yield, irrigation water use efficiency (IWUE), partial factor productivity of nitrogen (PFPN), and global warming potential (GWP). Treatments W3N2 (80%+155 kg/ha N) and W3N3 (80%+200 kg/ha N) achieved the highest yields, 11,883.51 and 11,436.82 kg/ha, respectively, whereas W2N1 (70%+110 kg/ha N) exhibited the best comprehensive performance, with a TCQ value of 0.65. Among the tested vegetation indices, the normalized difference vegetation index (NDVI) showed the strongest correlation with the basal crop coefficient, with an R²of 0.85. The NDVI -crop water stress index ( CWSI ) coupled model achieved the highest ET c estimation accuracy (R 2 = 0.89, RMSE = 0.39 mm/day), reducing the RMSE by 10.3% compared to the traditional, Multi-objective optimization revealed obvious trade-offs among high yield, water saving, high nitrogen efficiency, and low emissions. Scenario S5 was identified as the optimal solution, with an irrigation amount of 669.94 mm, a nitrogen rate of 117.48 kg/ha, a yield of 11,473.43 kg/ha, and the highest coordination degree of 0.86. These results demonstrate that coupling UAV multispectral and thermal infrared imagery with the FAO-56 model can effectively improve ETc estimation and provide reliable data support for water-nitrogen multi-objective optimization in cold-region rice production.
This study presents a multi-scale framework for reconstructing snow avalanche (SA) frequency and assessing vegetation structural responses in data-scarce mountain environments. The approach integrates dendrogeomorphological reconstructions, satellite-based spectral disturbance detection, and UAV-based Structure-from-Motion (SfM) photogrammetry, complemented by field data, and was applied to two avalanche paths in the Piatra Craiului Mountains (Southern Carpathians, Romania). Tree ring analyses allowed reconstruction of spatially explicit minimum avalanche chronologies for the 1980–2025 period. These reconstructions were combined with a DEM-based upslope algorithm to derive spatially variable avalanche return periods, revealing the highest frequencies in release and upper-track sectors and progressively longer return periods toward lower-track zones. Sentinel-2 imagery was used to assess the surface footprint of a reconstructed avalanche event in 2018. Among the tested spectral indices, the Moisture Stress Index (MSI) showed the most spatially coherent response, while the combined MSI-NDMI-NBR approach reduced index-specific noise. UAV-SfM photogrammetry supports high-resolution mapping of vegetation structure and surface states. Vegetation was classified using a machine-learning-based object-oriented approach (Random Forest) integrating spectral, geometric, structural, and textural parameters. The multi-parameter feature set yielded very high classification accuracy (Cohen’s Kappa ≈ 0.95). Across avalanche return-period gradients, both UAV-derived and field-based metrics showed a systematic associations between tree height and avalanche frequency, whereas tree age and stem diameter exhibited more variable, path-dependent responses. The proposed framework provides a transferable basis for linking avalanche disturbance regimes with vegetation structure and surface stability in mountain landscapes lacking long-term observational records.
Thermal imaging enables non-invasive assessment of canopy temperature, an essential indicator of plant stress, yet the lack of color cues and strong shadow interference make plant segmentation in thermal images difficult. Recent advances in foundation models have demonstrated improved performance and generalizability across applications, showing promise for domain-specific applications with limited annotated datasets such as plant segmentation in thermal images. This study investigates large multimodal models (LMMs) for thermal image segmentation in plant phenotyping. Building upon the SAM-CLIP framework, we design a unified pipeline spanning zero-shot inference, few-shot and low-shot fine-tuning, and active learning to maximize accuracy with minimal supervision. Evaluations on two thermal datasets, LadyBird Brassica and UGA Brassica, demonstrate robust performance after minimal adaptation across both datasets and superior performance compared with baselines, achieving mIoU D values of 97.54% on the LadyBird dataset and 76.94 % on the UGA dataset. We also release the resulting thermal segmentation annotations to support community benchmarking and reproducible research, highlighting the potential of LMMs to enable scalable, high-quality dataset construction for field phenotyping. The released datasets can be found at: https://cornell.box.com/s/dh69xf84464yrc1vlws92l1tflx7qa89
Reproduction assets foundThe authors publicly released the paper-specific thermal segmentation annotations (20,538 LadyBird masks and 37,790 UGA masks) via a Cornell Box link stated in the abstract, results, and data availability statement. No author analysis code or trained model checkpoints are explicitly released; the mmsegmentation GitHub/Dataset · publicwe generated and publicly released segmentation annotations for the complete LadyBird and UGA thermal image datasets using the best-performing SAM-CLIP model. Specifically, the final model obtained through the multi-round training process was used to generate 20,538 masks for the LadyBird dataset and 37,790 masks for the UGA dataset. Details of the generated annotations are provided in Supplementary Fig. S1 , and both annotated datasets are publicly available at: https://cornell.box.com/s/dh69xf84464yrc1vlws92l1tflx7qa89Open asset ↗lines:220-232Plant phenotyping relevance matchOpenAlex · checked 15 Sept 2026
Accurate estimation of pasture biomass is essential for determining cattle stocking rates and grazing durations. The objective of this study was to comparatively evaluate five sensor-based systems for estimating aboveground Bermudagrass (Cynodon dactylon) biomass and identify the leading sensing approach for continued development and broader validation. The five systems included Structure-from-Motion (SfM), Ultrasound Sensor and Ski (US-Ski), Inertial Measurement Unit and Ski (IMU-Ski), Inertial Measurement Unit and Roller (IMU-Roller), and Depth Camera (DC). These systems were deployed on unmanned aerial and ground vehicles to measure crop height under identical field conditions. Regression models relating measured crop height to wet biomass yield (WBY) were developed as a common calibration framework for statistically comparing sensor performance. These empirical allometric equations were intended to support comparative benchmarking of the sensing systems and were not developed as final operational biomass prediction models for immediate field deployment. The influence of vegetation coverage on yield predictions generated by the crop height-based equations was also examined. The results indicated that the IMU-Ski system demonstrated the strongest overall comparative performance (R2 = 0.97; SeY = 1112 kg-wet/ha), followed by the DC system (R2 = 0.97; SeY = 1132 kg-wet/ha). Based on its overall benchmarking performance, including calibration accuracy, residual error and simplicity, the IMU-Ski system was identified as the leading sensing approach for continued development and broader validation among the five evaluated methods. The results also indicated that addition of vegetation coverage into the crop height-based regression models did not significantly improve prediction accuracy under the experimental conditions evaluated.
Abstract In peanut ( Arachis hypogaea L.), plant stand establishment, seedling vigor, and canopy growth are key determinants of crop performance; yet traditional ground‐based assessment methods can be destructive, labor‐intensive, and limited in throughput. This study evaluated the potential of vegetation metrics derived from unmanned aerial vehicle (UAV)‐based red‐green‐blue (RGB) and multispectral (MS) imagery for high‐throughput, nondestructive assessment of plant stand establishment, seedling vigor, and light interception in peanut. Six runner‐type peanut cultivars were evaluated in 2024 and seven in 2025, with each cultivar represented by two seed size classes (small and large), to generate variation in these traits. Within‐row vegetation discontinuity‐based plant stand ratings for estimating plant stand count ( R 2 = 0.81–0.90), together with canopy coverage for assessing seedling biomass ( R 2 = 0.77–0.82) and light interception ( R 2 = 0.96–0.98), were the best‐performing vegetation metrics. These vegetation metrics provided similar or greater cultivar separation compared with ground‐based measurements. In contrast, several vegetation indices exhibited strong correlations with ground‐based measurements but provided inconsistent cultivar rankings and statistical groupings. MS imagery outperformed RGB imagery for plant stand and seedling biomass assessment. Overall, these results demonstrate that UAV‐derived canopy metrics provide reliable, high‐throughput tools for early‐ to mid‐season crop assessment and offer scalable alternatives to traditional ground‐based approaches for agronomic, crop physiological, and plant breeding research.
Accurate tree volume estimation is central to forest management and carbon accounting. Allometric equations are widely used but limited in transferability across species, regions, and environmental conditions. Mobile Laser Scanning (MLS) offers a promising alternative through direct measurement of tree geometry; however, the influence of tree shape on MLS accuracy remains poorly understood. This study evaluated MLS-derived estimates of stem diameters, total tree height, and merchantable stem volume against destructive reference measurements from 176 trees spanning eight species (four hardwood, four softwood) in Wallonia, Belgium. A Zeb Horizon RT scanner was used; tree architectural descriptors extracted from the point cloud were tested for associations with measurement error. Across 7,824 stem diameter measurements, MLS achieved a mean error of 0.46 cm, with precision declining above 15 m. MLS-derived height outperformed Vertex IV clinometer measurements for hardwood species (RMSE% = 6.88 vs. 8.78) but performed slightly less well for softwoods (RMSE% = 7.36 vs. 6.14). QSM-based volume estimates systematically underestimated reference values, while taper-based reconstruction produced nearly unbiased estimates with an RMSE of 15.72%. Correlation analyses and PCA showed that tree architectural variables explained only a small fraction of MLS error variability. Diameter and height errors were largely independent of structural attributes, while volume errors showed moderate associations with tree size and crown density. These findings indicate that tree architecture is not a primary source of MLS measurement uncertainty. Future MLS-based forest inventory efforts should prioritize acquisition and processing optimization, as scanning conditions and forest structure appear more influential than tree shape.
India’s economy is heavily reliant on agriculture, with a diverse array of crops grown on vast tracts of land. Fruit cultivation, especially papaya, has become more popular in recent years because of its high nutritional and financial value. To increase yield, maximize resource use and lessen reliance on chemical pesticides, modern techniques like protected cultivation and hydroponics are being used more and more. Fruit crops grown in controlled or semi-controlled environments are still susceptible to nutrient imbalances despite these developments, which can have a substantial impact on plant health, fruit quality and overall productivity. Papaya leaf nutrient deficiencies frequently show up in the early stages of growth and can result in poor fruit development and decreased yield if they are not detected in time. To support early diagnosis and better crop management, the current study focuses on creating an effective method for identifying nutrient deficiencies in papaya leaves using a deep learning (DL) framework based on transfer learning (TL). In this nutrient and micronutrient deficiency study and field work observation during year 2024 to 2026 with different climate and weather conditions done in order to tackle a new but related classification task, in the context of plant health assessment, several well-established architectures including InceptionV3, VGG19, DenseNet and Xception have been widely explored for leaf image analysis. Studies commonly utilize publicly available datasets, such as papaya leaf image repositories hosted on platforms like IEEE DataPort, to fine-tune these models for efficient feature extraction and accurate identification of nutrient and micronutrient deficiency patterns. This body of work demonstrates the growing role of transferring convolutional neural network (CNN) models in advancing automated crop monitoring and decision support systems.
Yang H, Zeng L, Gou X, Shao Y, Song X, Zhang S, Qian X, Zhang Z, Wang X, Li H, Peng B, Anyanwu JN, Zheng Z, Han L, Zhou L, Bulut M, Song P, Yang W, Xiao Y, Li W, Dai M, Qiu F, Zhang J, Wang B, Fernie AR, Wang H, Zhao H, Li X, Yan J, Guo T.
Improving nitrogen use efficiency (NUE) is essential for sustainable agriculture, yet conventionally measured plant characteristics have limited value as NUE proxies. Here we show that artificial intelligence (AI) can uncover previously unrecognized phenotypic variation associated with NUE, revealing genetic variation that is largely missed by conventional phenotypes. We trained a convolutional neural network (CNN) on 25,080 maize images to learn features that distinguish how plants respond to low- and high-N conditions, achieving 96.7% accuracy. The learned features were defined as deep phenotypes. Compared with conventional phenotypes, deep phenotypes showed greater phenotypic variation and higher heritability, enabling the identification of 523 significant loci compared with 21 for conventional phenotypes. We next investigated candidate genes underlying these loci and used these findings to interpret the learned features. Lower CNN layers primarily reflected visual patterns overlapping with conventional phenotypes, whereas deeper layers encoded additional features associated with N-responsive genetic variation. To validate candidate genes identified by the AI framework, we functionally characterized Liguleless2 (LG2), a basic-leucine zipper (bZIP) transcription factor, and demonstrated that lg2 mutants exhibit enhanced root architecture and increased N uptake efficiency. Field trials of 200 hybrids across diverse N environments further supported the AI findings, with each beneficial allele increasing ear weight by an average of 18 g per plot under low-N conditions. These results show how integrating AI and biology can uncover biologically relevant variation underlying complex traits such as NUE and enhance the interpretability of AI models.
Conventional methods of plant distance and volume measurements are limited by low efficiency, limited spatial coverage, and high measurement error. LiDAR and RGB-D imaging offer cost-effective, precise, and non-destructive techniques for plant distance and volume measurements. This study aimed to measure cabbage height, volume, and distance using LiDAR and RGB-D imaging. The sensors were mounted on a 1.6 kW electric field scouting platform (EFSP) for data collection. Point cloud (PCD) data were collected using LiDAR, whereas data processing, visualization, and measurements were done using commercial software and open-source programming scripts. A total of 20 cabbage plants were analyzed. LiDAR data processing included data frame screening, outlier removal, denoising, voxelization, and generation of 3D PCD density maps. Depth image processing included importing raw data and metadata shaping using intrinsic camera parameters, visualization, extraction of depth points, and pixel-level measurements of distances and volume. RGB image processing involved image conversion, segmentation, normalization, binary masking, mask cleaning, region extraction of cabbages, separation of ROI and preparation of contours, Delaunay triangulation and convex hull preparation, ROI overlay, bounding box preparation, sharing boundary between two boxes, conversion to pixel distances, and for visualization, plant height, volume measurements, and center to center distance measurement for measuring the plant distance. LiDAR demonstrated higher measurement accuracy for cabbage plant height, circumferential volume (geometric canopy volume), and plant distance, followed by RGB-D imaging, while RGB imagery showed comparatively lower performance under the study field conditions. Overall, LiDAR and RGB-D imaging provided reliable and non-destructive approaches for cabbage geometric characterization under field conditions, although accurately capturing complex plant geometry remains challenging. Positive and negative values of bias represent the over- and under-estimated results, respectively. Future studies should include larger and more diverse plant datasets exhibiting diversified size, shape, and geometric structure to further improve the robustness and general applicability of the proposed sensing approaches.
Corn is a staple crop of global significance; however, foliar diseases may lead to 30–60% yield loss if not detected at an early stage. Conventional visual inspection is time-consuming, subjective, and difficult to scale for smallholder farmers worldwide. To overcome these issues, we present Corn Transfer Learning Network (CTL-Net), an end-to-end hybrid deep learning model for corn leaf disease identification. CTL-Net, which combines Inception-ResNet-v2 as the backbone and MobileNetV3 as a feature extractor in parallel convolutional streams, simultaneously learns diverse scales of texture information from low-level textures, mid-level structural patterns, and high-level disease semantics of RGB leaf images. Adaptive feature fusion is formulated through learnable weighting coefficients and bi-directional spatial–channel attention mechanisms, which further enhance feature discriminability and robustness. The proposed approach is tested on an extensive dataset of 12,456 images from 10 corn diseases, including Northern Leaf Blight, Common Rust, Gray Leaf Spot, and Cercospora Leaf Spot, acquired under controlled and real-field conditions. CTL-Net attains the highest classification accuracy of 99.42%, outperforming DenseNet121 (97.92%), EfficientNetB3 (97.35%), and general stacking models (97.89%). Robustness experiments demonstrate the effectiveness of the proposed method against illumination variations, additive noise, and partial occlusions. CTL-Net enables real-time inference with a latency of 42 ms on an NVIDIA RTX 3090 GPU. Gradient-weighted Class Activation Mapping++ (Grad-CAM++)-based interpretability analysis results in a mean Intersection over Union (IoU) of 87.6% with expert-annotated disease regions. Five-fold cross-validation, ablation studies, and statistical significance testing (p
Reproduction assets foundThe paper states its final curated corn leaf disease dataset (12,456 images, 10 classes) is publicly available via the authors' GitHub repository vishruthkp/maizedataset (reference [30] and Data Availability statement). The Kaggle PlantVillage and Corn or Maize Leaf Disease datasets are cited source inputs, not paper-Dataset · public"Prediction of Crop Yield using Machine Learning," International Available: https://github.com/vishruthkp/maizedataset.Open asset ↗vishruthkp/maizedatasetpdf-page:7 lines:1-53Plant phenotyping relevance matchOpenAlex · Crossref · Europe PMC · checked 5 Sept 2026
Raquel Shany Betancur-Choquehuayta · Esther Emiliana Molina-Gonzales · Dony Javier Flores-Subeleta · Raymundo O. Gutiérrez-Rosales · Alberto Anculle-Arenas · Natty Wilma Llasaca-Calizaya · José Luis Bustamante-Muñoz · Mayela Elizabeth Mayta-Anco
Within the context of climate change, quinoa ( Chenopodium quinoa Willd.) is a climate-resilient crop with high nutritional value. The effects of deficit irrigation on quinoa growth and physiological performance under arid conditions remain insufficiently understood. This study evaluated ten quinoa genotypes (two commercial varieties and eight accessions) under two irrigation regimes to identify traits and spectral indices associated with water-stress tolerance. We combined manual phenotyping of agromorphological and physiological traits with multispectral and spectroradiometer measurements to calculate 35 vegetation indices across 13 and 5 dates, respectively. Deficit irrigation reduced plant height (18%), specific leaf area (8%), yield (43%), harvest index (26%), relative water content (7%), and dry matter accumulation (36%), while relative chlorophyll content (SPAD, Soil Plant Analysis Development) and stomatal density increased by 16% and 13%, respectively; accession ACC_23 exhibited the highest water-use efficiency (5.9 g kg -1 ). A univariate analysis of 35 vegetation indices across 13 dates showed that: Health Index(HIV), Normalized Green-Red Difference Index (NGRD), Red-Green Ratio (RG) and Plant Senescence Reflectance Index (PSRI), were the most sensitive, detecting significant differences between irrigation treatments in up to 32 of the 130 possible genotype-by-date comparisons. Integrating remote sensing into crop phenotyping represented a significant methodological improvement by enhancing phenotyping efficiency, improving detection of deficit irrigation effects, and facilitating identification of tolerant quinoa genotypes for arid production systems.
Grass seed crops are susceptible to yellow dwarf viruses transmitted by aphids. The Willamette Valley in Oregon, United States, is the leading producer of cool-season grass seed crops globally, and industry reports have attributed seed yield loss and shortened stand longevity to aphid-transmitted yellow dwarf viruses. Genetic resources are needed for effective and sustainable management of this pest, specifically the Rhopalosiphum padi–PAV pathosystem, in grass seed production to reduce foliar insecticide applications and maintain optimum seed yield potential. High-throughput phenotyping methods are needed to screen grass seed cultivars to identify resistant traits for traditional breeding programs. An automated video tracking procedure was optimized to evaluate host plant resistance in cool-season grass seed crops to R. padi–PAV with live plants and viruliferous and nonviruliferous aphid populations. Feeding behavior recorded with automated video tracking was strongly correlated with “ground-truthed” observations by human observers. Partial resistance (antixenosis and antibiosis) and tolerance traits were detected in select perennial ryegrass and tall fescue cultivars evaluated with traditional phenotyping methods in a greenhouse setting and with high-throughput phenotyping using automated video tracking in the laboratory. Across grass cultivars, nonviruliferous aphids had greater fitness and preference for noninfected grass plants compared with viruliferous aphids. Automated video tracking can be used as a high-throughput phenotyping method for continued evaluation of host plant resistance in grasses grown for seed production, as well as to identify resistant genotypes in other grass crops susceptible to aphid–yellow dwarf virus virus–vector systems.
leaf development was tracked dynamically from high-throughput phenotyping image series using the SAM-2 deep learning model, enabling automated segmentation and tracking of individual leaves across hundreds of plants from different accessions grown under different water and nitrogen conditions. From these time series, leaf-level growth dynamics were reconstructed and used to calculate key developmental traits, including phyllochron and relative expansion rates. These measurements were further integrated into the GreenLab model to infer plant-scale parameters such as radiation use efficiency (RUE) and leaf demand parameters. Statistical analyses revealed significant genotype, treatment, and interaction effects on several traits, with stable genotype rankings but contrasted sensitivities to environmental conditions. While whole-rosette growth captured strong and consistent responses, individual leaf contributions to global growth remained limited, highlighting the integrative nature of rosette-scale dynamics. Model-derived parameters provided additional insight into plant strategies, showing that genotypes differ in both the duration and timing of leaf demand, with generally stable patterns across environments. In parallel, QTL mapping was performed using leaf-level measurements and phyllochron traits. This analysis identified shared and trait-specific genetic loci, including loci associated with leaf emergence dynamics that were not detected using traditional rosette-scale integrative traits. Overall, this study demonstrates that combining deep learning-based image analysis with mechanistic modeling enables large-scale and quantitative characterization of plant development, providing biologically interpretable traits and new insights into their genetic and environmental determinism. It highlights the potential of combining high-throughput image-based phenotyping, deep learning, and mechanistic modeling to generate biologically interpretable traits and to assess genetic and environmental influences on these traits.
Reproduction assets foundThe paper explicitly states that the authors' leaf segmentation software (AraLeaf_segmentation), the GreenLab Arabidopsis model (AraGreenlab), and the statistical analysis code (AraParameters_statistical_analyses) are open-source and publicly hosted on forge.inrae.fr, and that the data and results used in the paper areCode · publicThe leaf segmentation software (AraLeaf_segmentation) and the GreenLab model of Arabidopsis thaliana (AraGreenlab) are open-source software, and distributed under the GNU GPL v3 licence. The two software packages, and the code to run statistical analyses (AraParameters_statistical_analyses) are available at:https://forge.inrae.fr/phenoscope-public/plant_phenomics_suppmat.Open asset ↗phenoscope-public/plant_phenomics_suppmathtml-lines:419-444Code / dataset availability confirmedCrossref · Europe PMC · checked 5 Sept 2026
Abstract Crop diseases pose significant challenges to productivity in resource-constrained settings, often remaining undiagnosed when diagnostic tools and infrastructure are either non-existent or inadequate. Current crop disease diagnosis relies on manual inspection methods that are labor-intensive, prone to error, and incapable of delivering real-time or region-specific insights in the process. Such limitations call for developing advanced diagnostic systems that are scalable and efficient in resource-constrained settings. This research introduced a comprehensive multi-spectral imaging and machine learning framework that can easily revolutionize the disease diagnosis and management inside the low-resource farming communities. Built within its core is the 3D Spectral-Spatial Convolutional Neural Network (3D SSCNN) that extracts high-resolution spectral-spatial features from hyperspectral image cubes. The accuracy achieved is around ~ 95% within 0.3 s per sample. Fed-DiagNet has provided support for distributed training that enables scalability and also data privacy to enhance the accuracy of regional models at approximately 92% as well as reduces training by almost 40%. Temporal disease progression modeling is enabled by Temporal Progression LSTM that provides dynamic trends with 90% accuracy up to a horizon of 10 days. This means that in addition to integrating disparate data sources-including hyperspectral imagery, environmental data, and pest observations-MTAN achieves an almost ~ 93% stress identification accuracy. Lastly, an RL-FO system tailors its treatment recommendations to local conditions so as to optimize for yield improvement and cost-effectiveness. With the proposed system, diagnostic precision increases to ~ 94%, and it is manifested in real-time efficiency while supporting scalability with actionable insights to empower farmers to mitigate crop losses and augment food security across several scenarios.
Reproduction assets foundThe paper's Data Availability statement points to two public repositories containing the data analyzed: a Kaggle PlantVillage dataset and a GitHub hyperspectral datasets repository. No author code or models are explicitly deposited.Dataset · publicAll data analyzed during this study are available in the Kaggle and Github repository, in the links https://www.kaggle.com/datasets/rohithaaiswarya/plant-village and https://github.com/antmedellin/HyperspectralDatasets .Open asset ↗Kaggle · rohithaaiswarya/plant-villagelines:563-583Dataset · publicAll data analyzed during this study are available in the Kaggle and Github repository, in the links https://www.kaggle.com/datasets/rohithaaiswarya/plant-village and https://github.com/antmedellin/HyperspectralDatasets .Open asset ↗GitHub · antmedellin/HyperspectralDatasetslines:563-583Plant phenotyping relevance matchOpenAlex · Crossref · checked 15 Sept 2026
Field / plotMultimodalNeRF / 3D Gaussian SplattingLiDAR / point cloudRGB / grayscaleStereoFruitWhole plant / canopy / plot / fieldMorphology / geometry measurementObject detection
Binocular stereo vision is a low-cost and scalable 3D perception technology that shows strong potential in agricultural phenotyping and smart agriculture. By estimating depth from multi-view RGB images, it enables non-contact, high-precision sensing of crop structure, canopy morphology, growth dynamics, and livestock traits, providing essential support for digital and intelligent agricultural production. With recent advances in deep learning-based stereo matching, multimodal sensor fusion, and 3D reconstruction, its robustness and accuracy in complex field environments have been significantly improved. This paper systematically reviews recent progress in agricultural applications of binocular stereo vision, covering system architectures, traditional and deep learning-based stereo matching methods, point cloud reconstruction techniques, and emerging supervision strategies such as 3D Gaussian splatting. It further summarizes key applications, including high-throughput phenotyping, fruit localization and robotic harvesting, weed detection and precision spraying, autonomous navigation, and livestock body condition assessment, highlighting its role in multi-task agricultural perception systems. Finally, the paper discusses major challenges, including low-texture matching difficulty, occlusions in complex environments, cross-domain generalization, real-time lightweight deployment, and limited dataset availability. Future directions are outlined in foundation model-based visual perception, self- and weakly supervised learning, multimodal fusion, and edge-efficient model design, aiming to support large-scale deployment in smart agriculture.
Sorghum (Sorghum bicolor L. Moench) is a major cereal in water-limited environments. Its C4 carbon-concentrating pathway suppresses photorespiration and supports comparatively high photosynthetic and water-use efficiency at high temperature, although yield remains sensitive to the timing and intensity of drought. This systematic review critically evaluates how coordinated variation in phenology, canopy development, transpiration regulation, photosynthetic resilience and root-mediated water capture can be phenotyped for sorghum improvement. The review was conducted and reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 statement. Eligible primary studies examined sorghum drought physiology, sensing-based phenotyping, trait retrieval, root-associated water capture, or breeding applications. Following duplicate removal and title, abstract and full-text screening, 45 sorghum-specific studies were included. Owing to substantial heterogeneity in experimental design, drought treatment, sensing platform, target trait, and validation metric, evidence was synthesised narratively rather than by meta-analysis. We compare sorghum studies across Light Detection and Ranging (LiDAR), multi-spectral, hyperspectral, thermal, structural, and fluorescence sensing, with emphasis on reported accuracy, transferability and physiological interpretation. We then examine how PROSAIL (PROSPECT coupled with Scattering by Arbitrarily Inclined Leaves) and SCOPE (Soil Canopy Observation, Photochemistry and Energy Fluxes) can be constrained for sorghum canopies and combined with machine learning. The central contribution is a sorghum-specific framework that distinguishes directly observed or model-retrieved canopy traits from indirect root-function predictions requiring ground validation. The synthesis identifies practical routes for measuring functional stay-green, high-vapour-pressure-deficit responses and post-anthesis water capture, while defining priorities for cross-environment validation and breeding deployment.
LiDAR (Light Detection and Ranging) and RGB-D camera imaging have emerged as essential tools in agricultural applications, particularly for plant size and distance measurements, enabling non-destructive, cost-effective, and precise estimation. The objective of this study was to measure the plant canopy dimensions and distance between apples using commercial LiDAR, and an RGB-D camera with a speed sprayer platform was used to determine whether LiDAR provides a higher measurement accuracy under field conditions. Data were collected in an apple orchard in Muju, Republic of Korea. Commercial 3D LiDAR, a terminal box, an RGB-D camera, a microcontroller, a power supply, and individual display monitors were integrated into a customized data acquisition (DAQ) box for LiDAR point cloud (PCD), RGB, and depth imagery data collection. Commercial software was used for data acquisition, data conversion (pcap to PCD), segmentation of regions of interest (ROI), and pre-processing of data. PCD processing and measurement consisted of data frame selection, data conversion, outlier removal, downsampling, denoising, ground point removal by filtering, voxelization, and density map generation using an open access programming language script. Depth image processing included importing raw data, shaping metadata using intrinsic camera parameters, visualizing depth images, extracting depth points, and measuring the plant canopy at the pixel level. RGB image analysis involved grayscale conversion, thresholding, segmentation of ROI, contour preparation, noise removal, and binary masking for eliminating the background. Estimated results were compared to measured results. LiDAR measurements showed the closest agreement with the measured results for plant height, canopy volume, plant spacing, and row distance, outperforming both RGB and depth imaging. Under field conditions, plant spacing and row distance were estimated with accuracies of 97.5% and 94.7%, respectively, exhibiting higher measurement accuracies than RGB and depth imagery data results. Despite some discrepancies due to complex plant geometry and dynamic data collection, the results support data collection strategies critical for precision horticulture.
Crop straw-to-grain ratio (SGR) estimation underpins regional straw resource assessment, yet national inventories rely on fixed coefficients that ignore structured variation across variety, environment, and phenotype. We introduce a Structured Multi-Kernel Heteroscedastic Gaussian Process (GP) framework that models SGR variation through three additive kernels heuristically motivated by the genotype–environment–phenotype (G+E+P) framework—capturing variety-associated variation, spatially structured variation, and environmental and management covariates—and employs an input-dependent noise model for prediction-specific uncertainty quantification. To prevent information leakage, target encoding and feature scaling are recomputed within each cross-validation fold. Evaluated via internal leave-one-out cross-validation on 80 rice samples (42 varieties, six Chinese provinces), the model achieves R2=0.541 with a prediction interval coverage probability of 0.95. Ablation identifies variety-associated variation as the largest contributor among the modeled factors (ΔR2=−0.024) and the multi-kernel design, by incorporating variety-specific information, substantially improves upon a covariate-only RBF GP (ΔR2=0.103). On point-prediction accuracy, Gradient Boosting achieves R2=0.58, slightly ahead of the Heteroscedastic GP (R2=0.54), underscoring that the primary advantage of the GP lies in its input-dependent uncertainty quantification. However, leave-one-county-out validation yields R2≈0 (with σ escalating to 24.4), confirming that the model does not yet generalize to unsampled counties; all reported performance is therefore internal to the nine sampled counties. The framework couples an agronomically motivated additive kernel structure with input-dependent uncertainty quantification, offering a path toward uncertainty-aware prediction from small field datasets.
Accurate assessment of crop water status is critical for precision irrigation and sustainable water management in agriculture. This study develops a UAV-based thermal infrared inversion framework for high-resolution canopy temperature retrieval and irrigation decision support in tea plantations. The proposed approach integrates multi-frame image mosaicking, threshold-based canopy extraction, and a gray-temperature calibration model to generate spatially continuous canopy temperature maps. Crop water stress was quantified using the Crop Water Stress Index (CWSI), and its reliability was further evaluated by analyzing its relationship with stomatal conductance. The framework further estimates soil moisture status and irrigation requirements based on a threshold-based irrigation strategy. The results show that the linear gray-temperature calibration model achieved a maximum absolute error of less than 0.3 °C and that the calculated CWSI and estimated irrigation requirement were strongly correlated with measured stomatal conductance, with R 2 up to 0.91. The proposed method provides a practical technical workflow from UAV thermal imagery acquisition to canopy temperature retrieval and quantitative irrigation decision-making, demonstrating its potential for precision irrigation management in tea plantations.
This study investigates the automatic classification of crop phenological stages from low-altitude UAV RGB imagery. The dataset included 11,489 images of five crops: sunflower, rapeseed, soybean, wheat, and barley. The images were annotated using the Biologische Bundesanstalt, Bundessortenamt und Chemische Industrie (BBCH) scale, with labels corresponding to either single stages or stage ranges to reflect heterogeneous field conditions and transitional crop states. A pretrained ResNet18 model was adapted to the task using transfer learning. Training was conducted in two stages: first, the classification head was optimized while the backbone remained frozen; second, the entire network was fine-tuned. The model achieved strong internal test accuracy across all crops, with 100% test accuracy for rapeseed and barley, more than 99% for the remaining crops, and a mean accuracy of 99.73% under the studied survey conditions. The results also compare favorably with previously reported studies on UAV-based phenological classification. Overall, the findings support the potential of low-altitude UAV imagery and deep learning for localized phenological assessment of selected field zones in precision agriculture, while broader deployment requires validation across independent fields, seasons, regions, and survey conditions.
Cereal crops, including wheat and barley, are essential for global food security, but their productivity is strongly affected by nitrogen availability and water limitation. This study investigated the phenotypic responses of two commercially significant spring wheat cultivars, Videodur (DU) and Sensas (SW), and two spring barley cultivars, Tiroler Imperial (SG1) and Amidala (SG2), exposed to two nitrogen regimes, low nitrogen at 25 kg N/ha (N25) and high nitrogen at 130 kg N/ha (N130), under drought and well-watered conditions. Plants were monitored from the late vegetative stage through maturity under controlled multivariable climatic conditions similar to field settings. A high-throughput phenotyping workflow was applied, combining precision watering, RGB imaging, infrared thermography, and VNIR–SWIR hyperspectral imaging to quantify plant growth, projected digital biomass, plant temperature, water use efficiency, and spectral vegetation indices associated with pigment dynamics, water status, maturation, and senescence. The results revealed cultivar-specific responses to combined nitrogen and drought stress. Under drought conditions, the high nitrogen treatment (N130) increased plant temperature (Tplant) for barley (cv. SG1) and wheat (cv. SW) compared to N25, thereby accelerating early maturation. However, the decline in chlorophyll was not uniformly faster across all cultivars tested. The DU cultivar exhibited superior chlorophyll absorption and reflectance, indicating better drought adaptation compared to other tested species. The high nitrogen treatment (N130) reduced water use efficiency (WUE) in the SW and SG2 cultivars compared to N25, implying that these cultivars used more water. Enhanced nitrogen did not consistently improve water use efficiency but did accelerate the growth cycle. SG2 was particularly sensitive to drought, showing declines in vegetation indices, except for the Water Content Index, highlighting the need for precise water and nitrogen management. Overall, the integration of hyperspectral, thermal, RGB, and water use measurements enabled the identification of trait signatures linked to drought adaptation, nitrogen response, maturation, and senescence. These findings provide practical insights for optimizing nitrogen and irrigation management and for supporting breeding strategies aimed at improving cereal crop resilience under climate-change-associated stress conditions.
Monitoring continuous agricultural canopies is fundamentally limited by the geometric constraints and computational bottlenecks of traditional 3D reconstruction. This study presents a 3D volumetric phenotyping pipeline integrating semantic mask generation with 3D Gaussian Splatting (3DGS) to quantify greenhouse cucumber canopy architecture. To drive component-specific optimization, we evaluated custom-trained convolutional networks (YOLO11) against a zero-shot foundation model (SAM3), determining that SAM3 provided the necessary boundary precision for accurate spatial isolation. The optimized 3DGS model outperformed implicit NeRF baselines, preserving fine-scale morphological details at real-time rendering speeds ( ≈ 48 FPS). To enable actionable measurement, a uniform voxelization protocol was applied to the point cloud, successfully neutralizing algorithmic densification bias. This technical framework yielded highly accurate physical geometry, achieving a Root Mean Square Error (RMSE) of ≤ 0.59 cm against in situ leaf measurements. Transitioning to agronomic interpretation, the pipeline was deployed to quantify complex canopy architecture. It mathematically mapped structural congestion zones and provided a temporal validation of a standard pruning intervention, explicitly capturing the geometric increase in lower-canopy porosity and the upward translation of biomass. This framework provides a robust, scale-accurate tool for monitoring plant architecture and guiding dynamic canopy management.
Unmanned aerial vehicle (UAV) imagery can support plot-scale crop phenotyping, but spectral, RGB and structural predictors may contribute differently to different traits. We compared six predefined feature groups for predicting soybean SPAD and plant height (PH) in a 1.3 ha field experiment in Sanya, China. The field contained 6197 soybean planting plots, of which 234 had paired SPAD and PH measurements. Multispectral bands, vegetation indices (VIs), RGB descriptors and digital surface model (DSM) metrics were extracted from DJI Mavic 3 Multispectral imagery. Six regression algorithms were evaluated using random fivefold cross-validation, spatial block cross-validation and nested spatial cross-validation. Under random cross-validation, ExtraTrees with multispectral bands, VIs and RGB descriptors produced the numerically highest SPAD performance (R2 = 0.589; RMSE = 6.66), while BayesianRidge with multispectral bands, VIs and DSM metrics produced the highest PH performance (R2 = 0.760; RMSE = 7.14 cm). Nested spatial cross-validation yielded R2 = 0.473 and RMSE = 7.56 for SPAD and R2 = 0.690 and RMSE = 8.13 cm for PH. G4 was selected in four of the five outer folds for SPAD, although the selected algorithm varied, and G5 was selected in all five outer folds for PH. VIs improved prediction of both traits relative to the original bands. Adding RGB descriptors produced only a small and model-dependent improvement for SPAD, whereas adding DSM metrics produced a larger and more consistent improvement for PH. The complete feature set did not outperform G4 for SPAD or G5 for PH. The retained models were applied to all 6197 plots to map SPAD, PH and their field relative combinations. Because all of the validations used one field and one UAV acquisition date, the results describe performance within this experiment and do not establish transferability to other sites, years or growth stages.
Semi-arid rangelands support livelihoods and key ecosystem services, yet sustainable management depends on accurate and scalable monitoring of herbaceous aboveground biomass (AGB). Field-based measurements are spatially limited, while satellite-derived vegetation indices often perform poorly in complex savanna systems such as the Kalahari. Using unoccupied aerial vehicle (UAV) structure-from-motion (SfM) photogrammetry, we evaluate the ability of fine-scale canopy height and spectral reflectance metrics to predict herbaceous biomass across a grazing intensity gradient. Canopy height consistently predicted AGB across all grazing levels, whereas normalised difference vegetation index (NDVI) effects were weak and grazing-dependent. The UAV-derived canopy height showed strong relationships with total herbaceous AGB, explaining up to 72% of observed variation, whereas vegetation greenness measured using NDVI showed limited predictive power. In contrast, predicting biomass of foraging importance proved challenging, with UAV-derived structural and spectral metrics explaining only a small proportion of variation. Together, these findings highlight the value of UAV-derived structural measurements over traditional spectral indices for fine-scale rangeland monitoring in semi-arid systems, while underscoring the limitations of current UAV-based spectral and structural metrics for assessing forage value across species and sites.
Semi-arid rangelands support livelihoods and key ecosystem services, yet sustainable management depends on accurate and scalable monitoring of herbaceous aboveground biomass (AGB). Field-based measurements are spatially limited, while satellite-derived vegetation indices often perform poorly in complex savanna systems such as the Kalahari. Using unoccupied aerial vehicle (UAV) structure-from-motion (SfM) photogrammetry, we evaluate the ability of fine-scale canopy height and spectral reflectance metrics to predict herbaceous biomass across a grazing intensity gradient. Canopy height consistently predicted AGB across all grazing levels, whereas normalised difference vegetation index (NDVI) effects were weak and grazing-dependent. The UAV-derived canopy height showed strong relationships with total herbaceous AGB, explaining up to 72% of observed variation, whereas vegetation greenness measured using NDVI showed limited predictive power. In contrast, predicting biomass of foraging importance proved challenging, with UAV-derived structural and spectral metrics explaining only a small proportion of variation. Together, these findings highlight the value of UAV-derived structural measurements over traditional spectral indices for fine-scale rangeland monitoring in semi-arid systems, while underscoring the limitations of current UAV-based spectral and structural metrics for assessing forage value across species and sites.
Semi-arid rangelands support livelihoods and key ecosystem services, yet sustainable management depends on accurate and scalable monitoring of herbaceous aboveground biomass (AGB). Field-based measurements are spatially limited, while satellite-derived vegetation indices often perform poorly in complex savanna systems such as the Kalahari. Using unoccupied aerial vehicle (UAV) structure-from-motion (SfM) photogrammetry, we evaluate the ability of fine-scale canopy height and spectral reflectance metrics to predict herbaceous biomass across a grazing intensity gradient. Canopy height consistently predicted AGB across all grazing levels, whereas normalised difference vegetation index (NDVI) effects were weak and grazing-dependent. The UAV-derived canopy height showed strong relationships with total herbaceous AGB, explaining up to 72% of observed variation, whereas vegetation greenness measured using NDVI showed limited predictive power. In contrast, predicting biomass of foraging importance proved challenging, with UAV-derived structural and spectral metrics explaining only a small proportion of variation. Together, these findings highlight the value of UAV-derived structural measurements over traditional spectral indices for fine-scale rangeland monitoring in semi-arid systems, while underscoring the limitations of current UAV-based spectral and structural metrics for assessing forage value across species and sites.
Magnetic field (MF) technologies have been explored in agriculture since the 1930s, with research activity increasing markedly since 2016. However, they have not achieved mainstream adoption, partly because no MF-specific validated methodology exists for evaluating their effects under realistic field conditions. Unmanned Aerial Vehicle (UAV)-based multispectral sensing represents a potential pathway to address this limitation by providing spatially explicit, non-destructive estimates of key canopy physiological variables at field scale, thereby enabling, for the first time, the systematic evaluation and validation of MF treatment responses under open-field conditions. To realise this potential, however, a common evidential basis must first be established by identifying crop physiological variables that are both consistently modulated by MF treatments and reliably detectable by UAV remote sensing. This study addressed this challenge through a dual-stream evidence synthesis of 216 peer-reviewed publications, comprising 102 studies on MF treatments in agricultural crops and 114 studies on UAV-based multispectral monitoring. Evidence from both research domains was synthesised to identify physiological variables that are simultaneously responsive to MF treatments and detectable through UAV remote sensing. Five direct bridge variables were identified: chlorophyll content, nitrogen use efficiency (NUE)/nitrogen assimilation, above-ground biomass (AGB), leaf area index (LAI), and yield. Chlorophyll content emerged as the strongest bridge variable, combining consistent MF responsiveness with UAV estimation accuracies of up to R2 = 0.90. Based on these findings, a conceptual framework was developed linking MF treatments, UAV-derived vegetation indices, ground-truth measurements, and machine-learning approaches for field-scale validation. The review revealed a complete absence of integration between the two research domains within the reviewed corpus, despite their strong biological and methodological compatibility. The proposed framework is conceptual and remains to be experimentally validated; it provides the first operational pathway for evaluating MF technologies under realistic farming conditions and may support future research on sustainable and digitally enabled crop production systems.
Phosphorus (P) deficiency severely limits soybean ( Glycine max L.) productivity. This study proposed a three-stage screening framework to identify reliable traits and P-efficient genotypes. In Experiment I, percent tolerance to phosphorus deficiency (PTPD) was calculated for ten growth parameters across 98 genotypes under P-deficient and control conditions. Principal component analysis and comprehensive evaluation identified six key indicators in Experiment I, which were subsequently refined to five indicators through further analysis: SPAD at V3 and R1, photosynthetic rate at R1, shoot dry weight at R8, and seed number per plant at R8. Experiment II re-evaluated these traits using 12 contrasting genotypes under three P levels, identifying CN 15 as the most P-efficient and SN 22 as the most P-inefficient. Experiment III further revealed that CN 15 maintained superior PSII performance and exhibited a 26.2% increase in grain P-utilization efficiency under 0 µM KH 2 PO 4 treatment. This integrated framework offers a preliminary reference for screening P-efficient soybean genotypes under controlled conditions, pending field evaluation.
Accurate retrieval of crop structural and physiological traits from remote sensing data remains challenging due to limited field observations and poor cross-platform generalization of data-driven models. This study develops a physics-informed transfer learning framework to quantify the contributions of improving simulated data fidelity and increasing model complexity to retrieving winter wheat leaf area index (LAI) and canopy chlorophyll content (CCC) from hyperspectral observations. Two PROSAIL-D datasets with default and physically optimized leaf angle distributions were generated to represent different levels of simulation fidelity. Four dual-branch deep learning architectures (CNN, CNN–SE, CNN–Transformer, and CNN–SE–Transformer) integrating spectral bands and vegetation indices were pretrained on simulated datasets and transferred to real observations using progressive fine-tuning. Model performance was assessed using ground-based and unmanned aerial vehicle (UAV) hyperspectral datasets, and SHapley Additive exPlanations (SHAP) analysis was applied to interpret feature contributions. Results demonstrated that transfer learning substantially improved cross-domain generalization, while enhancing simulation fidelity provided greater performance gains than increasing network complexity. The CNN–Transformer model pretrained on physically optimized simulations achieved the highest accuracy and robustness for both LAI and CCC retrieval. At ground and UAV scales, it achieved LAI estimation accuracies of R2 = 0.55 (RMSE = 0.63) and R2 = 0.53 (RMSE = 0.62), respectively. For CCC estimation, the model obtained R2 = 0.59 at both scales, with RMSE values of 36.12 μg cm⁻2 and 37.56 μg cm⁻2 for ground and UAV observations, respectively. SHAP analysis indicated that physically optimized simulations shifted model attention toward physiologically relevant vegetation indices, whereas default simulations induced stronger dependence on unstable visible wavelengths. Physically informed simulation design combined with transfer learning effectively reduces simulation to reality discrepancies, whereas increasing deep model complexity alone provides limited improvement. The proposed framework offers an accurate, interpretable, and scalable solution for cross-platform crop trait retrieval from hyperspectral observations.
Norka M. Fortuny-Fernández · Emma A. Juárez-Acosta · Pablo Martínez · David García‐Callejas · Anne Damon · Natalia Y. Labrín-Sotomayor · Yuri J. Peña–Ramírez
Understanding the spatial structure of tree communities is fundamental for evaluating ecological interactions and management dynamics in agroforestry systems. However, the structural complexity and small spatial scale of traditional agroecosystems often limit the use of conventional remote sensing approaches. Recent advances in drone-based photogrammetry offer new opportunities to reconstruct the three-dimensional structure of vegetation at high spatial resolution and to quantify tree-level structural attributes. In this study, we applied aerial photogrammetry from unmanned aerial vehicles (UAVs) to characterize the spatial structure of agroforestry systems in traditional home gardens (THGs) in the Yucatan Peninsula, Mexico. The immediate neighborhood structure of the tree community of 20 THGs distributed along a south–north precipitation gradient was analyzed using two focal species as anchor references: Spondias purpurea and Annona muricata. High-resolution orthomosaics and three-dimensional point cloud models were generated to estimate structural attributes, including tree height, crown area, crown surface area, and canopy volume, which were combined with field measurements of diameter at breast height. Spatial indices describing aggregation, dominance, and neighborhood diversity were calculated to evaluate tree spatial organization and potential interaction patterns. The UAV-derived structural metrics revealed significant differences in canopy architecture across regions and between focal species. Regardless of the focal species, trees in the southern region exhibited greater height, crown diameter, and canopy volume than those in the northern region. Moreover, the spatial arrangement of tree communities also differed depending on which focal species was considered as the anchor, suggesting contrasting strategies of canopy dominance and spatial coexistence. Finally, our results validate the use of drone-based photogrammetry as an effective approach for capturing fine-scale spatial structure in complex agroforestry systems. By enabling detailed three-dimensional reconstruction of tree canopies, UAV remote sensing offers an affordable, simple approach to investigate neighborhood interactions, management effects, and structural dynamics in traditional agroecosystems that are difficult to assess using conventional field- or satellite-based methods.
Efficient nitrogen (N) management is essential for sustaining crop productivity while minimizing environmental impacts associated with nitrogen losses. However, the high spatial and temporal variability of soil nitrogen dynamics and crop nitrogen status makes field-scale monitoring challenging, while conventional soil and plant sampling methods are labor-intensive, destructive, and provide limited spatial coverage. Recent advances in remote sensing technologies and machine learning (ML) offer promising alternatives for high-throughput, non-destructive monitoring of crop nitrogen status and related nitrogen dynamics in agroecosystems. This review synthesizes current progress in the use of proximal and remote sensing platforms, including unmanned aerial vehicles (UAVs), satellites, and ground-based sensors for assessing crop nitrogen status and inferring soil nitrogen availability. We examine spectral, thermal, and structural indicators, together with emerging sensor-fusion and time-series approaches. We also evaluate ML algorithms, including emerging foundation model approaches, for estimating crop nitrogen status and inferring soil nitrogen indicators, highlighting their performance, limitations, and transferability across environments. Particular emphasis is placed on field-scale applications in heterogeneous and water-limited systems, where nitrogen-water interactions critically influence crop responses. Finally, we discuss current challenges, including data scarcity, model generalization, and operational constraints, and outline future directions toward integrated, real-time decision support systems for precision nitrogen management. Overall, this review provides a comprehensive framework for leveraging remote sensing and data-driven approaches to improve nitrogen monitoring and enhance nitrogen use efficiency in diverse cropping systems.
In vivo analysis of plant stress responses remains a major challenge in precision agriculture, limiting dynamic optimization of crop growth under variable environmental conditions. Fluorescence imaging enables nondestructive tracking of stress biomarkers, but its accuracy is compromised by the low abundance of endogenous signaling molecules, tissue autofluorescence interference, and limited signal penetration depth. Here, we develop a ratiometric near-infrared IIb fluorescent probe for sensitive monitoring of endogenous hydrogen sulfide (H 2 S). This nanoprobe integrates lanthanide nanoparticles with H 2 S-responsive molecular units, enabling enhanced tissue penetration and high-resolution imaging through an absorption competition-induced emission mechanism. Under abiotic stress conditions, the probe visualizes stress-induced fluctuations of H 2 S in living plants. This work establishes an H 2 S-centered strategy for plant stress visualization and provides a foundation for developing early diagnosis platforms.
Ground-level ozone (O 3 ) adversely affects rice physiology and is associated with yield reductions. This study developed a high-resolution assessment framework integrating multi-source satellite remote sensing with econometric methods to quantify the impacts of O 3 on rice production in China's primary rice-growing region-the Middle and Lower Reaches of the Yangtze River (MLYR)-from 2019 to 2023. We fused Sentinel-5P TROPOMI total ozone column (TOC) data, a harmonized multi-satellite solar-induced chlorophyll fluorescence (SIF) product (LHSIF), high-precision rice distribution maps, and ERA5 meteorological reanalysis data. In addition to SIF, we examined multiple vegetation indicators (chlorophyll content, leaf area index, and vegetation indices) to capture broad physiological responses. A bidirectional fixed-effects panel model was employed to control for spatiotemporal confounders, revealing a significant inhibitory effect of O 3 on photosynthesis (β = -1.334 × 10 -5 , p 3 concentrations would increase regional SIF by 36.36%, while a commensurate 10% reduction in annual exposure could elevate rice yields by approximately 8.4%. This spaceborne remote sensing approach provides a robust and transferable methodology for the precise regional monitoring of ozone stress and for informing targeted mitigation strategies to safeguard crop productivity.
The use of glufosinate-resistant GM soybean has expanded, raising concerns about resistant weed development and unintended transgene flow. To support monitoring for timely management, we propose an early, non-destructive identification method using spectral images acquired from whole soybean plants after glufosinate treatment. We evaluated the potential of spectral imaging, using RGB, infrared (IR) thermal, and chlorophyll fluorescence (CF) sensors, for early detection of glufosinate resistance in soybean. In the dose-response test, the key spectral indices including NDI, temperature difference, F v /F m , and NPQ distinguished between resistant and susceptible soybeans within 4 to 24 hours after treatment (HAT). IR thermal and CF imaging showed higher sensitivity in identifying resistance than RGB imaging by detecting spectral responses associated with physiological changes before visual symptoms appeared. Validation test with a single dose treatment of glufosinate reconfirmed that image analysis by both the naked eye and machine learning (ML) can discriminate between resistant and susceptible soybeans in a single day after glufosinate treatment. ML-based classification using IR thermal index achieved 100% accuracy as early as 6 HAT and the classification by the naked eye using IR thermal images showed 96.6% accuracy at 24 HAT. These results suggest that plant imaging enables early and non-destructive identification of herbicide-resistant individuals by detecting early spectral changes to herbicide treatment. These findings support its use as a potential alternative to conventional diagnostic methods for detecting individuals containing transgenes in herbicide-resistant GM soybean cultivation for future applications in herbicide-resistant weed monitoring.
Purpose Conventional plant disease detection is time-consuming and prone to human error. The purpose of this study is to propose an edge artificial intelligence (AI)-based deep learning framework for plant disease detection under real-field conditions. The model integrates convolutional neural networks (CNNs) with a Sliding Window Mean Absolute Deviation (SWMAD) preprocessing technique to address illumination variability and complex background conditions. Design/methodology/approach A dual-layer CNN model is developed for plant disease classification using HSV segmentation, flood-fill segmentation and SWMAD preprocessing to handle real-field variations. The model is trained on PlantVillage and real farm images and deployed via TensorFlow Lite for real-time offline detection. TensorFlow Lite is used to enable efficient on-device inference for deployment in resource-constrained environments. Findings Benchmark experiments conducted on the PlantVillage dataset achieved a classification accuracy of 99.91% under controlled conditions. On the hybrid dataset comprising PlantVillage and real-field images, the optimized CNN framework achieved a validation accuracy of 95.01% following extensive evaluation of optimizers, layer architectures and worker configurations. The integration of the proposed SWMAD preprocessing technique into the finalized architecture further improved the validation accuracy to 97.38%, demonstrating enhanced robustness and classification performance under practical agricultural conditions in the final deployed model. Originality/value The originality of this study lies in several novel contributions. First, we introduce an SWMAD-based preprocessing technique, which enhances local statistical variations in leaf images by capturing pixel-level deviations from neighborhood intensity means. Unlike, conventional preprocessing methods, SWMAD is specifically designed to handle real-field challenges such as illumination variation, noise and complex backgrounds. The improvement in validation accuracy demonstrates the effectiveness of the proposed approach in capturing more discriminative features compared to existing techniques, thereby improving overall model robustness and reliability in practical agricultural environments.
This paper explores how deep learning methods can be used to monitor the health of crops and identify diseases, particularly for the apple crop. As the need for food security and sustainable farming methods increases, there is a strong demand for early detection of crop diseases. We have used a Convolutional Neural Network (CNN), based on the model of VGG16 architecture, since the model is known to be effective in image classification. The dataset contained 7771 training images for to enhance machines deep learning regarding plant diseases. Following that, validation image collection of 1747 images divided into four health conditions of the apple crops. In addition, the model was evaluated using 196 new images as a final test. To enhance the model capacity for recognizing diseases of real leaves rather than just memorize exact training pictures, data augmentation was used with ImageDataGenerator of TensorFlow. This means the training images were zoomed, rotated, and shifted to enable the machine detects more variations. Ten epochs of training were performed to measure the model accuracy. The results indicated that the model obtained significant improvement in training and validation accuracy from 56.43% to 78.12% and 92.94 to 96.93%, respectively. Most impressively, the final test dataset, which contained completely new images, scored an accuracy rate of 98%. The results indicate that in the architecture field the application of deep learning methodologies is effective, suggesting that automated detection of diseases by using sophisticated image analysis manages crop diseases identification efficiently. Combination of these methods successfully creates avenues for novel research to built real-time systems of crop disease monitoring, which help farmers increase their productions and farm their lands sustainably.
Phenotyping high-biomass perennial crops is laborious and the rate of genetic gain in conventional perennial crop breeding programs is typically low. So, it is especially important to identify methods that produce efficiency gains in the breeding process. Miscanthus is a C4 perennial grass with favorable characteristics for producing biomass as a feedstock for biofuels and diverse bio-based products. Increasing biomass yield will increase profitability and environmental benefits, so it is a key target for Miscanthus breeding. In addition, the identification of well-adapted genotypes across a wide range of environmental conditions requires the establishment of multi-environment trials (METs). Sparse testing is a genomic prediction-based strategy that reduces the phenotyping costs in METs by selecting a subset of genotypes to evaluate in a subset of environments and then predicts the performance of the unobserved genotype-environment combinations. A Miscanthus sacchariflorus (MSA) population comprising 336 genotypes observed across three environments was analyzed implementing sparse testing designs. Three prediction models considering main effects (environments, genotypes, genomic) and interaction effects (genotype-by-environment; G×E interaction) were implemented for forecasting dry biomass yield (YDY), total culm (TCM), average internode length (AIL), and culm node number (CNN). Multiple calibration sets based on different compositions and sizes were considered to evaluate performance in terms of the predictive ability (PA) and the mean square error (MSE) for a fixed testing set size. The training set size ranged from 52 to 112 to predict a fixed set of 224 unobserved genotypes across all three environments. The results showed that the model accounting for G×E interaction consistently presented the highest PA and the lowest MSE: for CNN (PA: ~0.77, MSE: ~0.5) and YDY (PA: ~0.70, MSE: ~1.3) while for TCM and AIL these ranged from ~0.28 to 0.41 and ~1.3 to 4.3, respectively. Overall, varying training sets and allocation strategies did not affect PA and MSE, with 52 non-overlapping and 0 overlapping genotypes per environment as the optimal cost-effective allocation framework. This suggests that implementing sparse testing designs could significantly reduce phenotyping costs by fivefold, without compromising PA in breeding programs for perennial crops such as Miscanthus .
Reproduction assets foundThe paper's data availability statement points to a public figshare deposit (DOI 10.6084/m9.figshare.31796794) containing the datasets analyzed in this Miscanthus sparse-testing genomic prediction study, including the phenotypic and genotypic data used for the models.Dataset · publicThe datasets analyzed for this study can be found in the figshare repository at https://doi.org/10.6084/m9.figshare.31796794 .Open asset ↗figshare · 10.6084/m9.figshare.31796794lines:603-621Plant phenotyping relevance matchCrossref · checked 15 Sept 2026
Published4 Aug 2026Journal of Agriculture and Ecology Research InternationalCited by 0 · OpenAlex ↗
Crop stress develops through a sequence that begins with molecular and biophysical perturbation, progresses through physiological dysfunction, and only later becomes visually apparent. Precision agriculture therefore requires sensors that can shorten the interval between stress onset and actionable diagnosis while preserving spatial context. This critical narrative review examines the complementary roles of nanosensors, plant-wearable and implantable electronics, proximal sensing, unmanned aerial vehicles, satellite remote sensing, and geographic information systems in early crop-stress detection. Literature published from 2000 to 1 June 2026 was selected through live searches of accessible scholarly indexes, DOI registries, publisher records, institutional repositories, and citation networks, with foundational studies retained where necessary. The evidence shows that nano-enabled interfaces can measure early biochemical, ionic, volatile, electrical, and microclimatic signals at high temporal resolution, whereas geospatial technologies reveal the distribution, persistence, and management relevance of stress across canopies and fields. Optical nanotube sensors, surface-enhanced Raman probes, electrochemical microneedles, ion-selective wearables, and flexible leaf sensors have demonstrated biologically meaningful signals before visible symptoms in controlled or pilot field settings. Yet most remain constrained by sparse sampling, crop-specific calibration, bio-interface effects, power and communication burdens, uncertain durability, and limited agronomic validation. Geospatial methods are operationally more mature, particularly thermal and multispectral imaging for water stress and hyperspectral imaging for disease and nutrient-related changes, but they often infer stress through non-specific proxies that are confounded by canopy structure, atmosphere, soil background, phenology, and co-occurring stresses. The strongest future architecture is therefore not a contest between nanoscale and landscape-scale sensing. It is a multiscale system in which physiologically specific plant sensors anchor and interpret spatial imagery, while remote sensing directs where high-specificity measurements and interventions are most valuable. Progress depends on prospective field trials, reference measurements, uncertainty-aware data fusion, interoperability, lifecycle safety assessment, and decision thresholds linked to economic and agronomic outcomes.
Smallholder tomato farmers in Northern Nigeria correctly identify major fungal diseases only about 41% of the time using unaided visual inspection, contributing to fungicide misapplication and avoidable yield loss. This study evaluated the field impact of SmartfarmerApp, a smartphone-based diagnostic application built on a validated convolutional neural network, on farmer disease identification accuracy. A pre-test/post-test controlled design allocated 240 tomato farmers across eight Local Government Areas in Kano and Kaduna States to an intervention group (n = 120, received the application) or a control group (n = 120, continued with conventional information sources), using computer-generated random allocation stratified by location and gender.
High perishability of vegetables associated with rapid physiological deterioration and microbial spoilage results in 30-50% post-harvest losses globally. In earlier days, post-harvest diseases were detected by visual inspection, microbial culturing, and molecular assays. These destructive methods are time-consuming and unsuitable for real-time monitoring. Volatile organic compounds (VOCs) have emerged as promising non-destructive biomarkers enhanced for early detection of quality deterioration and pathogen attack, often before visible symptoms appear. This review provides a thorough overview of current knowledge on VOC emissions in postharvest vegetables with their biosynthetic origins, classification, and roles in different kinds of stress responses and host microbe interactions. VOC alterations during spoilage and disease progression are systematically evaluated, highlighting vegetable group-specific patterns and quantitative dynamics of key biomarkers emitted naturally and due to mechanical and microbial spoilage. GC-MS, GC-IMS, PTR-MS, electronic noses, and biosensors are advanced analytical techniques that are critically compared with emphasis on their integration with machine learning for classification accuracy. Despite this significant progress, variability across cultivars and storage conditions, overlap between host- and pathogen-derived volatile metabolites, and an enduring gap between laboratory findings and commercial applications are major challenges that cannot be ignored. The development of real-time monitoring systems, vegetable-specific VOC databases, and integration with smart storage infrastructure powered by the Internet of Things and artificial intelligence must be prioritized in the future.
Junxiong Zhou · Xuechen Li · Chonghao Qiu · Lang Qiao · Xiaowei Jia · Qi Yang · Chishan Zhang · Leikun Yin · Nanshan You · Vipin Kumar · David Mulla · Ce Yang · Zhenong Jin · Licheng Liu
Accurate 3D crop monitoring underpins data-driven precision agriculture by enabling field-scale analysis of plant structure, growth dynamics, and management response. Modern 3D reconstruction methods perform strongly on generic benchmarks, but rendered appearance may not translate into metrically and agronomically useful geometry in crop fields. We introduce UAV3DCrop, a public benchmark of repeated multi-angle unmanned aerial vehicle (UAV) crop surveys. It contains 88,830 RGB images at $5280 \times 3956$ pixels, with a ground sampling distance of 3.6-5.8 mm, from 91 scenes spanning corn, soybean, wheat, and oat. Track A evaluates seven scene-optimized methods -- Neural Radiance Field (NeRF) and 3D Gaussian Splatting (3DGS) variants -- on held-out views, photogrammetry-referenced depth, and canopy-height recovery. Track B tests four pretrained feed-forward models on zero-shot camera-pose and geometry estimation. The scene-optimized methods rank differently across the three targets: Splatfacto-big leads appearance, whereas Scaffold-GS leads depth and is statistically tied with Splatfacto for canopy height. Among feed-forward models, MapAnything leads on seven of the eight metrics, while the remaining models vary more across crops and fail severely on absolute scale in a way that alignment conceals. Repeated acquisitions reveal further sensitivities that differ by output type and by model, associated with position within the acquisition sequence and with tie-point multiplicity. Current 3D reconstruction methods are therefore not yet interchangeable for agronomic use: no single method wins on appearance, geometry, and canopy height at once, and only one of four feed-forward models recovers usable metric scale. The dataset is publicly available at https://link-dev.github.io/UAV3DCrop/
Reproduction assets foundThe paper introduces UAV3DCrop, a public benchmark of repeated multi-angle UAV crop surveys (88,830 RGB images, 91 scenes, four crops) with refined poses, photogrammetric depth references, and linked canopy-height and effective-LAI field measurements. The dataset is explicitly stated to be publicly available under CC BDataset · publiche acquisition sequence and with tie-point multiplicity. Current 3D reconstruction methods are therefore not yet interchangeable for agronomic use: no single method wins on appearance, geometry, and canopy height at once, and only one of four feed-forward models recovers usable metric scale. The dataset is publicly available at https://link-dev.github.io/UAV3DCrop/ .
Keywords:
UAV imagery; agricultural datasets; crop-field reconstruction; neural radiance fields;
Gaussian splatting; feed-forward geometry.
1 IntroductionOpen asset ↗UAV3DCroplines:1-90Plant phenotyping relevance matchOpenAlex · Crossref · Europe PMC · checked 5 Sept 2026
Accurate identification and effective prediction of the maize tasseling stage are of great significance for guiding precision field management and ensuring stable crop yields. Conventional manual observation methods suffer from high labor intensity, poor timeliness, and strong subjectivity. In this study, based on an unmanned aerial vehicle (UAV) remote sensing platform, LiDAR point cloud data and RGB imagery were simultaneously acquired to construct digital surface models (DSMs) and digital terrain models (DTMs). Multi-dimensional statistical features were extracted to establish a high-precision plant height estimation method applicable to the entire growth cycle of maize. On this basis, the Logistic growth curve function was introduced to fit the dynamic changes in plant height, enabling the identification and early prediction of the maize tasseling stage based on the plant height growth curve. The research results indicate the following: (1) For maize plant height estimation, the LiDAR sensor outperforms RGB. The optimal accuracy is achieved by combining the 99th percentile of DSM with the minimum DTM, yielding a root mean square error (RMSE) of 0.17 m. (2) Based on the high-accuracy plant height time series, the point of inflection (POI) achieves the highest accuracy in tasseling stage identification, with an RMSE of 2.586 d under the reconstructed time series. (3) Prediction accuracy of the tasseling stage improves with increasing plant height threshold, and optimal performance is observed when the threshold is ≥1.6 m with a growth rate between 0.11 and 0.13. This study establishes a technical framework of "time-series perception-dynamic simulation-feature identification-early prediction", providing a scientific basis for automated monitoring and precision management of the maize tasseling stage. It holds significant theoretical and practical value for the advancement of smart agriculture and crop phenotyping research.
Accurate tracking and measurement of pollen dispersal in the atmosphere are essential for assessing cross-pollination risks, particularly in the case of genetically engineered (GE) crops. We conducted a series of unique release-recapture field studies with GE switchgrass in Oliver Springs, TN, USA. Two hundred transgenic switchgrass plants (Panicum virgatum L. "Performer") were planted at the center of a clear-cut field, with one block of 100 plants expressing orange fluorescent protein (OFP) under a switchgrass ubiquitin promoter (PvUBI1) and another block of 100 plants expressing OFP driven by a maize pollen-specific promoter (Zm13). Pollen was sampled from the atmosphere using fixed (ground-based) and mobile (drone-based) sampling devices at different distances from the source field, with Lagrangian stochastic dispersal simulations run for sampling periods using high-resolution wind measurements. The pollen emission rate was estimated by combining simulated and measured pollen concentrations, and strong diurnal trends were observed. Diurnal emission rate trends were positively correlated with wind speed, temperature, and vapor pressure deficit, while negatively correlated with relative humidity. In low-wind meandering conditions, incorporating changing wind direction into the dispersal modeling improved pollen emission rate estimation and model-measurement comparisons. This study assesses the effectiveness of high- and low-volume pollen samplers in relation to source strength up to 1 km from the source, enhancing understanding of pollen measurement techniques. Additionally, it is a proof-of-concept for drone-based pollen sampling and GMO pollen tracking using fluorescence measurements. Results from our experiments have significant implications for cross-pollination risk assessment, prediction, and management of airborne allergens.
Reproduction assets foundThe paper's Data Availability statement deposits all sampling data, modeling code, and simulation results on the Virginia Tech Data Repository (DOI 10.7294/25733604), which is an allowed URL. This directly covers the paper's pollen concentration measurements and Lagrangian stochastic dispersal modeling. Other URLs (e.gDataset · publicAll sampling data, modeling code, and simulation results underlying this manuscript are made available on the Virginia Tech Data Repository at https://doi.org/10.7294/25733604 .Open asset ↗Virginia Tech Data Repository · 10.7294/25733604lines:201-219Plant phenotyping relevance matchOpenAlex · checked 14 Sept 2026
Plant height is a key 3D phenotypic trait for assessing crop growth, biomass accumulation, and lodging resistance. To overcome the practical limitations of conventional plant height measurement methods, this study proposes a novel vision foundation model-based framework named Depth for Plant Height (Depth4PH) for plant height estimation in agricultural scenes using low-cost monocular RGB imaging. As part of our contributions, a synthetic–real coupled multimodal dataset was constructed by integrating Blender virtual agricultural scenes (Blender VAS) with real field images. Building upon the existing Depth Anything V2 foundation model, we developed a novel module called Transfer-based Agricultural Metric Depth Anything V2 (TAM-Depth V2) for absolute metric depth estimation through parameter-efficient fine-tuning, depth decoder reconstruction, and joint loss optimization. Furthermore, we designed a novel multi-source prompt-based segmentation framework, MSP-SAM2, to generate positive and negative prompts for zero-shot crop instance segmentation. Finally, a new inverse physical plant height estimation algorithm, RANSAC-Per, was introduced to estimate plant height by combining truncated percentile statistics with local RANSAC micro-plane fitting, thereby reducing the effects of depth noise and field microtopographic variation. The result showed that TAM-Depth V2 achieved stable absolute depth estimation, with an RMSE of 0.1162 m and an AbsRel of 4.25%. Compared to the original box-prompted SAM 2, MSP-SAM2 achieved a 4.4% improvement in mIoU, reaching 91.6% and a recall of 93.2%. On a 350-plant multi-crop test set, Depth4PH achieved R 2 = 0.948, RMSE = 12.23 cm, and MAE = 8.82 cm, and MAPE =10.15%, with crop-specific RMSEs ranging from 3.99 cm (cucumber) to 20.73 cm (maize), significantly outperforming the traditional Global-MinMax baseline (which had an RMSE of 22.62 cm). These results indicate that Depth4PH provides a promising foundational pathway for high-throughput crop phenotyping. With future optimization for edge deployment, it holds significant potential to support high-throughput monitoring in precision agriculture.
Annotation scarcity, poor model generalization and lagged data processing remain key bottlenecks hindering the practical deployment of phenotyping robots. To address these issues, we developed a novel phenotyping robot capable of online 3D reconstruction and zero-shot segmentation directly on the edge. Diverging from conventional semantic SLAM, our core contribution is RT-ZSDR, a framework featuring two key methodological novelties. First, we introduce the ForeCut pipeline for target extraction, which innovatively fuses DINO features with 3D geometric spatial information, leveraging multi-view semantic-spatial consistency to achieve annotation-free, zero-shot dense segmentation and reconstruction. Second, we designed a hardware-coupled loop closure strategy utilizing the robotic arm's kinematic feedback as prior constraints to significantly improve loop closure recall. Supported by edge computing Jetson Orin NX, the tracking and segmentation process takes approximately 0.24 s per frame after an initialization period of 1.82 s. RT-ZSDR's phenotypic measurements demonstrated strong correlations with reference baseline in both laboratory settings (n=90, PlantEye measurements as reference baseline; R 2 =0.990, 0.939, 0.725, and 0.861 for plant height, projected leaf area, surface area, and volume) and practical greenhouse environments (n=48, manual measurements as reference baseline; R 2 =0.965, 0.862 for plant height and stem diameter). Additionally, evaluated against COLMAP benchmarks (n=24), the system achieved a mean 3D reconstruction F1-score of 0.816.
Pin-Jun Wan · Qi-Xin Yang · Yu-Biao Cai · Xin-Feng Wang · Ya-Xuan Wang · Feng-Xiang Lai · Qi Wei · Jia-Chun He · Wei-Xia Wang · Jin-Li Zhang · Qiang Fu
The brown planthopper ( Nilaparvata lugens ) is one of the most destructive pests of rice and poses a threat to yield stability and food security. Although host-plant resistance is the most sustainable strategy for BPH management, conventional resistance phenotyping remains labor-intensive, destructive, and poorly suited for large-scale breeding. Here, we combined hyperspectral reflectance profiling of 50 rice varieties with an interpretable machine learning framework to enable non-destructive prediction of resistance phenotypes. Using post-infestation spectral profiles, we established classification models that captured resistance states shaped by constitutive traits and inducible defense responses. Among 13 evaluated algorithms, a radial basis function support vector machine achieved the best performance on full-spectrum data within the sampled variety panel, with an average accuracy of 0.939 ± 0.015 and a maximum of 0.972. Predictive wavelengths were concentrated in the green, red-edge, and near-infrared regions, corresponding to variation in pigment regulation, canopy structure, and water status. Spectral and network analyses showed that resistant genotypes exhibited more complex but less stable spectral co-occurrence networks, consistent with physiological trade-offs associated with defense. We also tested whether resistance could be predicted before pest infestation. Pre-infestation spectra retained significant predictive power, with accuracies of 0.572 ± 0.021 for five-class classification and 0.667 ± 0.021 for binary classification, indicating that constitutive defense-associated physiological states are optically detectable before visible damage occurs. Together, our results show that hyperspectral reflectance encodes both inducible responses after infestation and constitutive defense baselines present beforehand. This work establishes a scalable, non-invasive phenotyping strategy for early resistance screening within evaluated germplasm panels, while future validation across independent and variety-level held-out populations will be required before broader deployment.
Accurate and efficient acquisition of seedling density and growth information is of great significance for guiding modern agricultural field management. Although drone imagery has been widely used in seedling monitoring, the inherent trade-off between operational efficiency and image resolution limits the effectiveness of remote sensing-based seedling detection. To address this challenge, this study proposes an integrated analytical method combining super-resolution reconstruction and object detection. The approach first employs the Real-ESRGAN model to enhance low-resolution image quality, then utilizes the YOLOv12 model to accurately localize cotton seedlings, and finally generates visualizations of seedling density and growth uniformity. Experimental results demonstrate that super-resolution reconstruction enhances the detection algorithm's capability for small targets, increasing the object detection precision by 3.3%. With the incorporation of super-resolution reconstruction, the seedling counting accuracy reaches 92.08%, representing a 46.15% improvement over the method without super-resolution, thereby effectively enhancing the algorithm's counting capability. Furthermore, this method achieves image detail equivalent to that obtained at 7.5 meters flight altitude while operating at 30 meters, reducing data acquisition time to 1/16 of the original requirement. In practical applications, the visualized results of seedling density and growth uniformity provide precise decision-making support for thinning, replanting, and differentiated field management. The proposed method is not only applicable to cotton but can also be extended to staple crops such as corn, wheat, and rice, with additional potential applications in forestry and ecological monitoring.
To address the challenges of insufficient sensor stability and poor consistency among multi-source data during crop phenotyping in unstructured field environments, this study develops a hardware-software co-optimization framework for wheat canopy sensing based on a four-wheel-drive, high-clearance phenotyping platform. A multi-sensor stabilization device integrated with an ESO-LQR control strategy suppresses pitch disturbances during motion. A ROS-based hierarchical framework coordinates LiDAR, cameras, and inertial sensors, while spatial calibration and timestamp-based software synchronization ensure spatiotemporal consistency. A tightly coupled LiDAR-IMU SLAM algorithm enables centimeter-level 3D reconstruction of farmland. To mitigate terrain effects, a two-stage ground point extraction method integrating verticality and spatial distribution features improves canopy height estimation. Field experiments demonstrate stable platform operation at 1.5 m s⁻¹ while maintain high efficiency and data quality, with a coverage efficiency of 95.2%, retained-point ratio of 85.3%, and an MTF of 0.35 for image clarity. Phenotypic evaluation shows strong agreement between predicted wheat plant height and manual measurements, with R² values of 0.898 and 0.729 and RMSE values of 1.30 cm and 1.64 cm at the jointing and grain-filling stages, respectively. Moreover, at the jointing stage, both the 2D green area index (GAI) and the 3D point-cloud-based canopy coverage exhibit strong consistency with ImageJ-derived results (R² = 0.873 and 0.910). These findings demonstrate that the proposed approach enables stable and efficient acquisition of crop phenotypic information in complex field environments, providing a solid technical foundation for digital field monitoring, data-driven crop management, and intelligent agricultural systems.
Optimizing environmental inputs for indoor crop production by conducting a traditional endpoint growth analysis requires significant time and resources. The most common scientific approach to assessing crop response involves the accumulation of dry mass at the end of a cropping cycle. A growth dynamics analysis also results in the accurate estimation of the crop response to the growth environment through periodic destructive sampling. Measuring crop gas exchange in the same environment in which it is grown offers a powerful alternative to accelerating the environmental optimization process, especially for vegetative crops. This work introduces Minitron III, a third-generation technology advancement capable of continuous gas-exchange monitoring from seed to harvest for small specialty crop stands. For proof of concept, 24 ‘Rouxai’ red oakleaf lettuce plants were grown from seed to harvest over a 25-day cropping cycle. Instantaneous differences in the carbon dioxide (CO 2 ) and water vapor (H 2 O V ) mole fraction between sample/reference lines flowing through/around cuvette/growth space were measured using a differential infrared gas analyzer, allowing determination of net photosynthesis based on a 0.41-m 2 cropping area. Crop stand net photosynthesis was detectable 7 days after sowing seeds, increasing gradually from 0.13 to 0.60 µmol·m −2 ·s −1 over the following week. The crop net photosynthesis rate increased robustly on a daily basis from 15 days after sowing seeds. While the net photosynthesis rate at the beginning of the photoperiod was 0.68 µmol·m −2 ·s −1 on day 15, it increased to 7.7 µmol·m −2 ·s −1 by day 25 after sowing seeds. Crop dark respiration was detectable from 17 days after sowing seeds and ranged from −0.3 to −0.9 µmol·m −2 ·s −1 . Minitron III has potential for rapid optimization of multiple environmental inputs for indoor production of specialty crops based on the near-real-time crop response to environmental inputs.
• Development of yellow color index (YCI) for yield estimation at flowering stage • Development of web-based interface (YCPM-UAV) for canola yield prediction using UAVs • Global application capability for UAVs datasets to predict canola yield using YCPM-UAV • Multi-sensor and multi-spectrum data fusion to find most suited indices for canola • Multiple and stepwise regression analysis for selection of most influencing VIs Canola ( Brassica napus L.) is a globally significant oilseed crop, yet accurate yield estimation remains challenging due to the complex and unique nature of the crop, especially at the flowering stage. Traditional field-based yield estimation methods are labor-intensive, time-consuming, and destructive, necessitating innovative approaches for early and non-destructive yield prediction. The main objective of the study is to develop a novel web-based platform, YCPM-UAV (Yellow Color Prediction Model using Unmanned Aerial Vehicles), for early and accurate canola yield estimation using high-resolution multi-sensor datasets acquired through low-altitude UAVs (LA-UAVs). To achieve this objective, a comprehensive two-year field study (2022-2024) was conducted across ten farmers’ fields in different geographical locations. Multisensor data (RGB, multispectral, and thermal) were acquired using UAVs at seven growth stages. Several vegetation indices (VIs), yellow color-based indices, and a thermal index were calculated. Linear, multiple, and stepwise regression analyses were performed to evaluate relationships of remote sensing indices with ground-truth yield data collected from 1200 sampling points. Multiple and stepwise regression analyses indicated that the newly developed Yellow Color Index (YCI) exhibited the strongest correlation with actual canola yield at the flowering stage across both years (Year 1: R 2 = 0.84, RMSE = 39.30 g m⁻²; Year 2: R² = 0.88, RMSE = 31.57 g m⁻²). Based on proposed predictive modeling, the YCPM-UAV web interface was developed, featuring automated data processing and spatial analysis with a testing accuracy of 88%. The YCPM-UAV platform provides farmers, researchers, and policymakers with a timely, user-friendly, and actionable decision-support tool for canola yield estimation at the field scale, contributing to improved crop management and food security. Future studies should incorporate additional canola varieties, irrigated and non-irrigated fields, and deep learning algorithms to further improve model robustness.
• First PRISMA-ScR mapping of 79 HSI-ML asymptomatic detection studies (42 species, 74 pathogens) • Controlled-to-field accuracy gap quantified: 91.4% vs. 86.3% (5.1 pp, p = 0.0163 ) • 56.8% of studies omit temporal sampling documentation (CV = 139%) • SWIR underutilization (11.8%) reflects economic, not scientific, barriers • DBVS proposed as standardized temporal metric for cross-study comparability Plant disease management requires non-invasive detection methods capable of identifying infections before visible symptom manifestation, thereby enabling timely intervention. Hyperspectral imaging combined with machine learning and deep learning (HSI-ML) achieves 90.2% classification accuracy in controlled environments for asymptomatic plant detection; however, systematic characterization of methodological practices across this rapidly expanding field remains absent. This PRISMA-ScR compliant scoping review mapped 79 peer-reviewed studies (2010–2025) encompassing 42 plant species and 74 pathogenic agents using a Population-Concept-Context framework. Visible-near-infrared (VNIR) systems dominated deployment (61.8%, n = 49 ), while short-wave infrared (SWIR) systems remained substantially underutilized (11.8%, n = 9 ) due primarily to economic rather than scientific constraints. Among 67 unique algorithms identified, machine learning methods accounted for 30.7% (SVM, random forests, and PLS-DA predominant), whereas deep learning represented 28.4% (2D-CNN, 3D-CNN, and hybrid architectures). Critical methodological gaps emerged: 56.8% of studies omitted temporal sampling documentation (detection latency range: 1–56 days post-inoculation; coefficient of variation = 139%). Platform-stratified analysis revealed controlled environments achieved 91.4% ± 6.6% classification accuracy ( n = 48 ) versus 86.3% ± 9.1% for field/UAV deployments ( n = 26 ), representing a significant 5.1 percentage-point performance decrease ( p = 0.0163 ). Detection accuracy exhibited a weak negative correlation with detection timing ( ρ = − 0.33 , p = 0.067 ), though this association did not reach conventional statistical significance. Methodological heterogeneity—rather than algorithmic limitations—constitutes the primary barrier to field operationalization. Adoption of Days Before Visible Symptoms (DBVS) as a standardized temporal metric could resolve an estimated 40–50% of cross-study variance currently attributed to inconsistent asymptomatic-phase definitions.
Accurate plant population estimation is critical for crop monitoring, yield prediction, and field management in precision agriculture. However, challenges such as complex backgrounds, overlapping plants, and the need for large-scale coverage make seedling counting from UAV imagery difficult. In this study, we propose a novel automated pipeline for maize seedling counting based on high-resolution UAV orthomosaic images. The pipeline begins with an automatic field-based plot extraction method that isolates individual planting regions without the need for manual intervention. A lightweight object detection model, YOLOv8n-CA, is then employed, incorporating Coordinate Attention to enhance feature localization while maintaining fast inference. To further accelerate the process, we introduce the ‘In-Range Sliding’ strategy, which limits inference to only planting regions, reducing computational overhead. Extensive experiments demonstrate the effectiveness of our approach, achieving a mean absolute error (MAE) of 0.587 and a coefficient of determination ( R 2 ) of 94.19 % at the plot level. Additionally, our system enables spatial feature extraction such as seedling spacing, supporting more refined agronomic analysis. This work provides an efficient and scalable solution for plant counting, offering valuable insights for large-scale, rapid post-processing agricultural monitoring.
Kai Zhou · Yuxin Jiao · Guangning Yu · Xin Wang · Weijun Zhang · Shaolong Zhu · Yuxiang Zhang · Wenyan Yang · Shujun Hu · Xinyi Wang · Yulin Yin · Ziqiu Wang · Xuecai Zhang · Zefeng Yang · Tao Liu · Chenwu Xu · Yang Xu
Plant height (PH) and aboveground biomass (AGB) are critical agronomic traits that determine the yield potential of maize ( Zea mays L . ). However, the application of genomic selection (GS) and genome-wide association studies (GWAS) in maize breeding is often hindered by the limitations of phenotypic data collection, which is typically characterized by low throughput and inadequate accuracy. To address this challenge, we employed an unmanned aerial vehicle (UAV) equipped with LiDAR and RGB cameras for high-throughput assessment of pH and AGB in a panel of 817 maize hybrids derived from 364 inbred lines over two growing seasons. Our results demonstrated that the integration of UAV-derived LiDAR point clouds with crop surface models (CSMs) enabled robust estimation of pH across multiple years ( R 2 > 0.90). Furthermore, a three-dimensional AGB estimation model was developed using UAV-derived PH and canopy coverage (CC), achieving high estimation accuracy ( R 2 > 0.83). Subsequently, the UAV-derived PH and AGB were utilized for GS and GWAS analyses. Replicated 10-fold cross-validation showed that the mean predictability was 0.504 for PH and 0.402 for AGB across eight commonly used GS models. Moreover, of the 66,066 potential crosses derived from the 364 inbred lines, the top 200 crosses selected for AGB showed up to twice the AGB of the bottom 200 crosses. Field validation demonstrated that the mean ear weight (EW) in the AGB top group was 39.0% higher than that in the bottom group. A total of 16 and 11 significant SNPs were identified by at least two GWAS methods for PH and AGB, respectively. Based on these SNPs, 81 candidate genes were functionally annotated, six of which were simultaneously associated with both traits. The candidate gene association analysis suggested that variations in the promoter region of ZmFLA9 may affect both traits. Overall, our study highlights the potential of UAV-based high-throughput phenotyping to accelerate maize genomic breeding by enabling rapid, precise, and large-scale trait assessment.
Accurate and non-destructive counting of rice seedlings is crucial for yield estimation and precision agriculture, yet remains challenging in UAV videos due to dense distribution and strong temporal appearance similarity. This study proposes an efficient tracking-based rice seedling counting framework that integrates an improved Yolov11n detector with a robust multi-object tracking strategy to achieve reliable video level counting. The proposed detector, termed DMNP-YOLO, enhances feature representation, localization robustness, and computational efficiency through Dynamic Snake Convolution, a multi-scale feature attention module, Shape-IoU combined with Normalized Wasserstein Distance, and BatchNorm scaling factor based structured channel pruning, resulting in reductions of 40.5% in Params and 15.2% in GFLOPs while achieving a precision of 0.901 and an mAP@0.5 of 0.921. Building upon accurate frame-level detections, a trajectory based counting mechanism is realized by embedding an Anchor–Angle–Distance association strategy into ByteTrack, which explicitly enforces geometric and temporal consistency across frames, significantly improving tracking stability in dense seedling scenes. As a result, Multi-Object Tracking Accuracy is increased by 5.3 percentage points, identity switches are reduced by 33.3%, and counting accuracy is improved by 3.7 percentage points. Extensive experiments demonstrate that the proposed tracking-based counting framework achieves a mean absolute error of 16.47, a mean absolute percentage error of 6.48%, and an R² of 0.95969. Field scale validation further confirms its practical applicability, achieving an overall rice seedling counting accuracy of 93.4% and demonstrating strong robustness in real world agricultural environments.
Lei Li · Jindong Liu · Guoliang Wan · Hongqing Wang · Mengjiao Yang · Shuaipeng Fei · Duoxia Wang · Yong Zhang · Xianchun Xia · Xin Ma · Yong He · Yonggui Xiao
The extent of canopy coverage (CC) prior to the booting stage is a useful indicator of environmental adaptation and may help anticipate key developmental events such as heading and flowering. We used UAV-based high-throughput phenotyping to monitor CC in a 262-line F 8 recombinant inbred line population grown under four irrigation-year environments across two seasons. To quantify CC dynamics, we fitted regression models using either days after sowing (DAS) or accumulated active temperature (AT). These models showed a high goodness-of-fit (overall coefficient of determination ( R 2 ) > 0.90 across environments; per-timepoint prediction R 2 = 0.60‒0.99 with root mean squared error (RMSE) = 0.00‒0.02) and were used to derive 31 CC-related traits. Principal component analysis (PCA) showed that the first two components explained 80.00% of total variance in CC traits, with PCA1 accounting for 48.88%‒62.51% of the variance for DAS-based traits and 50.26%‒54.93% for AT-based traits. PCA1 reflected early canopy vigor and rapid coverage increase, while PCA2 reflected canopy maintenance after jointing. Genotypes in the top 15% for PCA1 differed significantly in flowering time, heading time, and plant height from those in the bottom 15%. Cross-environment comparisons showed moderate to high consistency of CC traits (average correlation ( r ) = 0.41‒0.65 for DAS-standardized CC of 20‒160 DAS and 0.48‒0.67 for AT-standardized CC of 100‒1000 AT). Using CC features derived from both DAS and AT, we trained a random forest model to predict flowering time, highlighting the contribution of both temporal and thermal information to predictive accuracy. This model achieved an independent test set R 2 of 0.81 with an RMSE of 1.25 days within the current dataset. All 31 CC-derived traits were also used for QTL mapping. Inclusive composite interval mapping identified 156 QTL detection events across 15 chromosomes (0.80%‒27.70% phenotypic variance explained), which were consolidated into 26 QTL regions, including loci such as QCC.caas.7A (671.47‒680.09 Mb on 7A) and QCC.caas.5D2 (426.67‒459.42 Mb on 5D). Several of these regions co-localized with previously reported genes associated with tillering, flowering time, winter hardiness and plant height. These findings indicate that CC, characterized by DAS- and AT-based traits, is genetically tractable and predictive of flowering within the current dataset, supporting its potential as a phenology-related trait for wheat improvement. Further validation across broader environments, years, and genetic backgrounds will be needed before broader application.
Accurate, field-scale mapping of crop growth stages is critical for supply-sensitive vegetable production, where timely harvests require detailed phenological information. Consecutive growth stages often involve rapid and subtle morphological changes and are influenced by challenging open-field conditions, which frequently result in misclassification when stages are treated as independent, discrete categories. To address this issue, CropMap is proposed as a growth-stage mapping framework that integrates Hierarchical Semantic Segmentation Networks (HSSN) with multispectral unmanned aerial vehicles (UAVs) imagery. CropMap incorporates the structured biological progression of crop development into the learning objective through tree-based label constraints, allowing the model to recognize phenological continuity and reduce confusion between adjacent stages. The framework is evaluated on the publicly available National Information Society Agency of Korea (NIA) field crop growth-stage dataset, a large-scale, multi-institutional UAV dataset containing 337,665 multispectral patches across six hierarchically related growth stages of Chinese cabbage and radish, curated by the NIA. CropMap achieves a test-set mean Intersection over Union (mIoU) of 0.5382, representing a 5% relative improvement over the best-performing transformer baseline (SegFormer; mIoU = 0.5124). Performance varies across classes: background separation is strong (IoU = 0.9128) and the rosette stage is well distinguished (IoU = 0.6354), while the leaf expansion stage remains the primary challenge (IoU = 0.3541), reflecting the inherent difficulty of mapping this spectrally and morphologically transitional class. These findings indicate that hierarchy-aware learning reduces inter-stage confusion for most phenological classes, but transitional growth stages remain a significant limitation for field-scale deployment. The framework provides a foundation for stage-resolved crop monitoring to support harvest timing and supply forecasting in high-value vegetable systems.
Cover crops offer essential agroecosystem benefits, including reduced soil erosion, weed suppression, and improved soil health. Aboveground biomass (AGB) is a key indicator of these benefits; however, field-based quantification is often limited, which hinders effective cover crop management decisions. This study integrated unmanned aerial vehicle (UAV)-based multispectral imagery with machine learning (ML) models to estimate AGB in cover crops across two water-limited regions of Texas. Ground-truth and imagery data were collected over three years (2023–2025) for winter rye ( Secale cereale L.) in Lamesa and two years (2023–2024) for winter wheat ( Triticum aestivum L.) in Chillicothe under varying irrigation regimes. Five ML algorithms, random forest, support vector regression, extreme gradient boosting, partial least squares regression (PLSR), and artificial neural network (ANN), were evaluated across four individual and eleven feature fusion datasets. The ANN model consistently achieved the highest predictive accuracy, particularly when vegetation indices were combined with structural features (R² = 0.87, RMSE = 9.08 g m - ²), while PLSR showed the weakest performance. Grouped validation (leave-one-year-out, leave-one-species-out, and leave-one-treatment-out) revealed reduced model performance compared to random (70/30) splitting of pooled data, yet the ANN maintained moderate predictive ability, indicating reasonable generalizability across years, species, and management conditions. Shapley additive explanations (SHAP) revealed key predictors in the ANN model, including plant height, chlorophyll vegetation index, chlorophyll sensitive index, blue band reflectance, modified chlorophyll absorption in reflectance index, dissimilarity, correlation, and enhanced green vegetation index. These findings demonstrate the effectiveness of UAV-ML integration for accurate AGB estimation and highlight the potential for scalable, data-driven cover crop monitoring in water-limited environments and beyond.
Shoot apical meristem (SAM) homeostasis integrates environmental and genetic cues to regulate growth dynamics that drive biomass accumulation and crop yield; however, no robust, non-destructive, quantitative proxy has been established for modeling or monitoring SAM-homeostasis-associated dynamics. Here, we developed a novel robot-based 3D imaging system and a custom pot-chamber gas exchange system to non-destructively measure plant occupation volume (POV) and whole-plant photosynthetic rate in wild-type Arabidopsis plants and nine mutants with disrupted SAM homeostasis. We demonstrate that POV robustly captures 3D plant architecture, whereas whole-plant photosynthetic rate serves as a superior proxy for optimal growth dynamics and final biomass associated with SAM homeostasis, outperforming conventional traits such as leaf number, leaf size, total leaf area, and rosette diameter. The strong positive correlations among POV, whole plant photosynthesis, and biomass accumulation establish a powerful new framework for quantitative studies of SAM homeostasis and data-driven evaluation of plant architecture.
Reproduction assets foundThe paper's authors explicitly state that the Python source code for whole-plant leaf-area segmentation, 3D point cloud processing, POV calculation, and Mask3D-based segmentation is publicly available on GitHub at https://github.com/songqingfeng/AtPOVcalculator. This is a paper-specific, public, actionable analysis/PhDCode · publicThe Python source code for whole-plant leaf-area segmentation and calculation is publicly available on GitHub ( https://github.com/songqingfeng/AtPOVcalculator ).Open asset ↗songqingfeng/AtPOVcalculatorlines:224-233Plant phenotyping relevance matchCrossref · checked 15 Sept 2026
Biomass is a key trait in pasture plant breeding and agronomy, but measuring Dry Matter Yield or Fresh Weight across large numbers of samples is labour intensive and costly. Efficient biomass assessment systems must balance accuracy, speed, and cost, while ideally enabling non-destructive measurements. We developed a rapid, real-time, non-destructive, and remotely controlled LiDAR-based platform to estimate biomass in grass monocultures by measuring sward height at high spatial resolution. The system operates under ambient light conditions at a ground speed of 2.7 km per hour. It was evaluated in small-plot perennial ryegrass trials at two field sites in New Zealand, across two seasons at site A and one season at site B. At site A, correlations between LiDAR-derived height and fresh weight ranged from 0.33 to 0.74 across individual measurement cycles, with an overall multilevel R² of 0.72. At site B, the multilevel correlation increased to R² = 0.88. Weekly LiDAR scans at site B were used to estimate plot-level growth rates for 60 plots, demonstrating improved temporal resolution. Statistically significant differences in growth rate within regrowth cycles were detected among plots. The platform reliably differentiates perennial ryegrass plots based on biomass and offers higher temporal resolution than traditional methods.
Shannon Baker · Lei Zhao · Jose Landivar Scott · Jackie Rudd · Amir MH Ibrahim · Juan Landivar · Jinha Jung · Shuyu Liu · Kevin Nowka · Mahendra Bhandari
Multi-temporal data from unoccupied aerial systems (UAS) offer insights into growth parameters for winter wheat breeding decisions. Weekly UAS data were collected during the 2019 and 2020 growing seasons from dryland and irrigated nurseries at Bushland, Texas, within the Texas A&M Uniform Variety Trials. Canopy cover (CC) was extracted from orthomosaic images and modeled using a double-sigmoid function with a second-order derivative that captured genotypic variation in canopy growth and senescence with high coefficients of determination (R² > 0.99) and low root mean square error (RMSE) values ranging from 1.73 to 4.06. Analysis of variance (ANOVA) revealed highly significant genotypic effects (p<0.001) for yield, heading, and Green Leaf Area Duration (LAD) in all environments except 2020 dryland, where no significant differences among genotypes were detected. Extracted parameters showed positive correlations with agronomic traits, particularly under rainfed and stress-prone conditions. The end decrease stage (EDS) was correlated with grain yield (r=0.56 in 2019 dryland, and r=0.53 in 2020 irrigated, p<0.001), and the start decrease stage (SDS) was highly correlated with yield (r=0.53 in 2020 irrigated, p<0.001). The maximum decrease rate date (MDRD) was positively correlated with yield in 2019 dryland (r=0.55, p<0.001), while LAD had a correlation of r=0.56 in 2019 dryland and r=0.58 in 2020 irrigated (p<0.001). These findings demonstrate that double-sigmoid model provides a powerful, non-invasive framework for quantifying canopy development, senescence timing, and stress responses. By distinguishing genetics from environmental influences on canopy dynamics, this approach enhances selection accuracy and accelerates the development of stress-resilient winter wheat cultivars.
Accurate monitoring of cotton plant moisture content (PMC) is crucial for guiding irrigation practices. To address the limited capacity of single-source remote sensing data to characterize the water status of cotton plants, as well as the lack of quantitative reference values for suitable PMC levels at different growth stages, this study constructed a cotton PMC estimation model based on multimodal UAV remote sensing data. Furthermore, the suitable reference levels of PMC at different growth stages were investigated according to the response relationship between PMC and yield at each growth stage. Five soil moisture gradients were established, and at each growth stage, fresh and dry weights of cotton shoots were measured to calculate the PMC. A UAV platform equipped with multiple sensors was used to collect visible-light (RGB), multispectral (MS), and thermal infrared (TIR) images of the cotton canopy. Three feature selection methods were employed to identify moisture-sensitive parameters: Pearson correlation analysis, principal component analysis (PCA) for dimensionality reduction, and recursive feature elimination (RFE). Using the selected parameters, four machine learning algorithms, AdaBoost, random forest (RF), CatBoost, and k-nearest neighbors (KNN), were applied to construct and validate PMC estimation models. The suitable PMC levels at different growth stages were identified based on the response relationship between measured PMC and yield under different water gradients. The results showed that the RFE feature selection method identified eight water-sensitive parameters, and the CatBoost model integrating multimodal data performed best, with R² and RMSE reaching 0.807 and 0.033%, respectively, on the test set, providing a reliable method for high-resolution spatial mapping of field-scale PMC. On this basis, the response of yield to PMC was analyzed, revealing that when PMC was maintained at 83.8%, 85.9%, 79.3%, 78.0%, and 67.7% at the bud, initial flowering, peak flowering, peak boll-setting, and boll opening stages, respectively, the theoretical maximum yield of 6579–6667 kg/hm² could be achieved. This study realized high-precision remote sensing monitoring of PMC and further explored the appropriate moisture content thresholds for different growth stages, providing a quantitative reference for precision water regulation in cotton fields.
【Objective】Chlorophyll fluorescence is a physiological indicator reflecting crop photosynthesis and water stress. Non-destructively monitoring the changes in chlorophyll fluorescence under water stress is critical for improving irrigation management. This paper explores the applicability of canopy hyperspectral reflectance for elucidating the response of rice canopy chlorophyll fluorescence to water stress.【Method】The experiment was conducted in pots and the measurements were taken during the booting stage of rice. Three water treatments were set, including continuous flooding irrigation (CK), mild drought (MS) and severe drought (HS). Canopy hyperspectral reflectance and chlorophyll fluorescence were synchronously measured using a high-throughput phenotyping platform, from which we analysed the responses of chlorophyll fluorescence traits to soil water change. Prediction models were developed to estimate chlorophyll fluorescence traits using partial least squares regression (PLSR) and backpropagation neural network (BPNN), based on characteristic spectral bands.【Result】①The chlorophyll fluorescence traits Fv/Fm, Y(II), qL and Y(NPQ) varied with water stress, with significant changes observed 3-4 days after cessation of irrigation, and detectable variation identified up to day 6 after terminating irrigation. On day 6 after irrigation cessation, the HS treatment reduced Fv/Fm, Y(II) and qL by 41.3%, 46.9% and 53.1%, respectively, whereas increased Y(NPQ) by 117.5% compared with CK. ②Savitzky-Golay smoothing and multiplicative scatter correction (MSC) preprocessing effectively reduced the scattering effects on canopy hyperspectral data induced by structural variation. The characteristic spectral bands selected from the hyperspectral data were mainly distributed in the blue (400-500 nm), red and near-infrared regions. ③Compared with PLSR, the BPNN was more effective in capturing the nonlinear relationships between hyperspectral data and chlorophyll fluorescence traits. The BPNN was most accurate for estimating Y(NPQ) and qL, with the associated R2 values being 0.867 and 0.845, respectively, and less accurate for estimating Fv/Fm.【Conclusion】Canopy hyperspectral data can be used to estimate rice chlorophyll fluorescence traits. This approach provides a rapid, cost-effective, and non-destructive method for monitoring crop physiological responses to water stress.
Quantifying the canopy growth dynamics and light interception capacity under different management practices laid the physiological foundation for potato yield formation. However, the traditional manual measurement methods are labour-intensive, time-consuming, and incapable of capturing time-series dynamics. To address this, we proposed a novel high-throughput strategy that integrates UAV-based RGB imaging with a piecewise physiological model. Furthermore, how Nitrogen(N)-Potassium(K) interaction affects the temporal canopy growth dynamics, light interception, and tuber yield was determined. The results indicated that: (1) Among the 11 secondary indices extracted from the canopy growth dynamic curves, the interaction of N and K had the greatest effect on the maximum canopy duration and the total canopy growth curve integral. The direct path coefficients of N and K inputs on these two parameters were 0.847 and 0.805, and 0.234 and 0.148, respectively. (2) There was a strong linear relationship between the integral area under the curve (S∫) and the total plant dry weight, with R² at 0.90 in 2023-2024. A simplified net photosynthetically active radiation utilisation assessment framework that achieved high accuracy with minimal parameter was built. (3) Prolonging the maximum canopy continuous coverage time is the main way to improve potato yield. The overall effect of N input on yield was significantly higher than that of K fertiliser, with a total effect value of 1.428. Optimising the N-K interaction improves nutrient precision and light interception. The integration of UAV remote sensing and the crop physiological-ecological model enables the tracking of potato canopy dynamics, which is helpful for optimising management practices to improve potato yield.
Accurate prediction of foxtail millet yield is essential for effective field management and high-throughput breeding. Despite advances in UAV-based yield prediction for major crops, existing studies predominantly rely on single-temporal features (SFs) extracted at noon, overlooking significant diurnal dynamic signals that characterize crop responses to water stress. To address this research gap, we propose a novel approach utilizing diurnal cross-temporal features (CFs) derived from UAV-based multispectral and thermal imagery to enhance yield prediction accuracy under different irrigation regimes. During the flowering and grain-filling stages, UAV images were acquired across eight time slots (T1–T8) within a single day to capture the complete diurnal trajectory of canopy physiological responses. SFs were extracted at each time slot, and CFs were derived through summation, averaging, and range operations across multiple slots. A systematic four-step workflow was developed to determine the optimal UAV flight frequency and timing by balancing prediction accuracy with operational costs. Three ensemble learning algorithms (Random Forest (RF), Adaptive Boosting (AdaBoost), and Extreme Gradient Boosting (XGBoost)) were evaluated using multiple feature sets incorporating SFs, CFs, and their integration. Results demonstrated that CFs more comprehensively captured dynamic crop responses to water stress than SFs. Canopy features from afternoon combinations generally exhibited stronger yield correlations than morning combinations. The [T5, T8] combination was identified as optimal, providing a practical balance between prediction accuracy and operational cost. Model comparison revealed that RF exhibited greater robustness across different water treatments, whereas AdaBoost achieved higher accuracy on the test set. Feature importance analysis confirmed the dominance of CFs, with ∑VSWI ranking first across both models and growth stages. This study provides a systematic framework for utilizing diurnal dynamic signals in crop yield prediction, offering new methodological insights for precision agriculture and high-throughput phenotyping of foxtail millet and other dryland crops.
Citrus fruit cracking causes substantial yield and economic losses, yet its relationship with plant water status (PWS) and irrigation management remains insufficiently characterized. Unlike previous UAV-based irrigation studies that focused on water-stress detection or yield estimation, this study introduces a dynamic, physiology-based framework that links temporal PWS trajectories during key phenological stages to fruit-cracking risk at the individual-tree scale. UAV-based multispectral, thermal, and LiDAR data, combined with field physiological measurements and machine-learning models, were evaluated in an irrigation management experiment in an ‘Ori’ mandarin orchard (Israel) across three contrasting growing seasons (2023–2025). Several irrigation treatments with different irrigation timings and water inputs were applied during the growing season to evaluate their effects on temporal PWS dynamics and fruit cracking. Trunk growth (TG), stem water potential (SWP), stomatal conductance (SC), and plant area index (PAI) were measured throughout the two seasons and estimated using Random Forest models (R 2 > 0.783). These indicators were subsequently used to predict yield and fruit cracking with high accuracy (yield: R² = 0.896; cracking: R² = 0.845). Cracking was lowest in 2023 (∼3%), with ∼25% lower irrigation, suggesting reduced irrigation may reduce cracking risk. Higher cracking in 2024 (∼14%, vs ∼8% in 2025) coincided with intense heat events. Mid-season SWP and SC were strongly associated with yield formation and cracking patterns. These findings demonstrate that monitoring temporal PWS dynamics can support precision irrigation management by identifying high-risk zones and enabling irrigation strategies that stabilize PWS, reduce the incidence of cracking, and improve yield under variable climatic conditions.
Accurate estimation of forest aboveground biomass (AGB) is critical for carbon accounting, ecosystem monitoring, and climate change mitigation. While remote sensing offers broad spatial coverage, standard deterministic models often struggle to capture the complex relationship between spectral information and structural forest attributes, and they frequently lack reliable quantification of uncertainty. This study presents a novel probabilistic framework that integrates convolutional neural networks (CNNs) and transformer architectures for AGB estimation using a multi-sensor fusion of Sentinel-1/2 imagery. The framework employs a customized, adaptive version of the Swin Transformer-reconfigured specifically for regression tasks and streamlined to a single-stage architecture-and implements a Gaussian Mixture Modeling (GMM) approach to ensemble model outputs. This framework enables the decomposition of predictive uncertainty into epistemic (model-related) and aleatoric (data-inherent) components. Evaluated across heterogeneous forest regions in Montana and Idaho, the CNN-transformer integration achieved superior performance, with R 2 =0.83, RMSE=19.16 Mg/ha, and MAE=13.37 Mg/ha, outperforming standalone CNNs (R 2 =0.81) and vision transformers (R 2 =0.80). Transfer learning and fine-tuning experiments demonstrated high model robustness, improving prediction accuracy in independent test regions by 39% in RMSE. Spatially explicit analysis revealed that while CNNs provide stable local feature extraction, the Swin Transformer’s self-attention mechanism significantly mitigates errors on topographically complex slopes by leveraging non-local spectral cues to compensate for shadowing and geometric distortions. The results highlight that an ensemble approach leveraging the complementary strengths of CNNs and customized transformers provides the transparency and precision required for operational carbon monitoring and high-resolution, trustworthy biomass mapping.
The precise identification of unsound soybean seeds is a critical step in deep soybean processing and seed selection. The accuracy of this identification directly influences the quality of subsequent processed products, as well as the germination rate and yield of soybean crops. This study proposes a nondestructive identification method for unsound soybean seeds based on hyperspectral imaging (HSI), Gramian Angular Field (GAF), and a Dual-Channel Residual-Squeeze-and-Excitation Network with GAF Fusion (DC-RSEN-GF). According to common damage types, soybeans were categorized into six classes: sound seeds, thermal-damaged seeds, insect-damaged seeds, broken seeds, spotted seeds, and moldy seeds. Spectral data from these six soybean categories were acquired using a hyperspectral camera and transformed into two-dimensional GAF images. The DC-RSEN-GF network integrates one-dimensional spectral data with two-dimensional GAF images. After preprocessing with Savitzky-Golay (SG) smoothing, high-precision classification was achieved through residual blocks, an attention mechanism (using SENet), and feature fusion. Compared to five benchmark models-Extremely Randomized Trees (ERT), Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), VGG19, and ResNet18-the DC-RSEN-GF model achieved superior performance, with accuracy, precision, specificity, and F1-scores of 96.36%, 96.43%, 97.92%, and 96.36%, respectively. The accuracy, precision, and F1-scores are all superior to traditional machine learning and existing deep learning models, demonstrating better classification capabilities. In addition, t-distributed Stochastic Neighbor Embedding (t-SNE) was employed for visual analysis of soybean spectra, further validating the reliability of the DC-RSEN-GF model. The proposed detection method, based on HSI and DC-RSEN-GF, enables accurate and nondestructive identification of unsound soybean seeds and holds significant potential for practical application.
Japanese agriculture faces pressing challenges, including a declining and aging farming population and the need to adapt to climate change. To address these issues, Smart Agriculture is being introduced to improve production efficiency. Among these, unmanned aerial vehicles (UAVs) have gained attention for their ability to rapidly monitor entire fields. We proposed a machine learning-based crop growth diagnosis system that generates spatiotemporal data for multiple vegetation indices (VIs) using the quartile method and diagnoses crop growth based on patterns of change in these values. The experimental site consisted of five paddy fields within an 80 m × 50 m plot in Iwate Prefecture, Japan, equipped with weather and water sensors. Ground-truth data (overall length, culm length, panicle number, and stem number) were collected approximately one week before harvest. UAV monitoring was conducted four times using a multispectral camera, and growth analysis was performed with six VIs. Correlation analysis revealed a positive relationship between crop growth and the daily average water level during the drainage period, and a negative relationship with the daily temperature range in mid-June. A combined cluster-label representation, constructed from clustering results of all VIs for each mesh, enabled integrated analysis and visualization of multi-index patterns. Grid size optimization showed no significant differences in correlation trends between 1 m × 1 m and 5 m × 5 m resolutions. For non-crop area removal, a comparison of three image segmentation methods demonstrated that the Otsu Method achieved the highest performance. Finally, to facilitate practical use in the field, we prototyped a report interface for the diagnosis system. Future work will focus on developing a comprehensive field diagnosis system to clarify field environments, with the aim of addressing fragmentation and enclaves in Japanese farms.
In-season fine-scale (i.e., within-field experiment plot scale) crop grain yield (GY) prediction is critical for optimizing inputs, minimizing environmental impacts, and supporting sustainable food production. Traditional approaches, such as field surveys, are often costly and inefficient over large areas. As an alternative, remote sensing combined with crop simulation models (CSMs) has been increasingly applied for in-season GY prediction. This study investigates the potential of integrating Uncrewed Aircraft Systems (UAS)-based remote sensing data, deep learning, and CSMs to predict maize and soybean GY using a data assimilation approach. UAS multispectral imagery was collected, along with field-measured maize above-ground biomass (AGB) and soybean leaf area index (LAI) during the 2022 and 2023 growing seasons at experimental fields in Brookings, South Dakota. Maize AGB was measured at two growth stages, while soybean LAI was collected across four stages. One-dimensional convolutional neural networks (1D-CNNs) were used to estimate maize AGB and soybean LAI from canopy spectral, textural, and structural features derived from UAS imagery. These UAS and deep learning–derived crop traits were assimilated into DSSAT-Maize and DSSAT-Soybean models to optimize parameters, and the optimized models were subsequently used to predict GY. For maize, the DSSAT-Maize model achieved an R² of 0.62, an RMSE of 717.8 kg ha⁻¹, and an rRMSE of 6.7% for GY prediction. For soybean, the DSSAT-Soybean model achieved an R² of 0.81, an RMSE of 207.3 kg ha⁻¹, and an rRMSE of 4.9%. Overall, these results highlight the potential of combining high-resolution UAS data and deep learning–derived crop traits within a CSM framework through data assimilation, enabling fine-scale, in-season yield predictions and supporting precise agricultural management.
• DL model trained on MS data achieved the highest accuracy with an R 2 of 75.29% • Linear Regression Coefficient-Based feature selection with SVM and PCA with Linear Regression significantly improved model performance. • NIR and red-edge bands in the MS dataset consistently outperformed the RGB dataset • More represented rice variety (Sona) achieved a strong R 2 of 80.21% on MS data Accurate crop yield prediction is critical for agricultural planning, food security assessment, and farm-level decision-making. In Nepal, however, rice yield estimation is still predominantly based on traditional approaches, where local agricultural extension offices collect field-level observations that are subsequently aggregated at district, provincial, and national scales, often limiting spatial detail and timeliness. This study aims to develop a field-scale rice yield estimation framework by integrating Unmanned Aerial Vehicle (UAV)-derived remote sensing data with machine learning (ML) and deep learning (DL) techniques. High-resolution multispectral (MS) and RGB UAV imagery were used to evaluate the influence of Vegetation Indices (VIs), including HUE and VNDVI from RGB data and RGBVI and Simple Ratio (SR) from MS data, along with plant characteristics and farm management practices (e.g., application of Zyme and Zinc Potash) on rice yield. The predictive performance of Support Vector Machines (SVM), Linear Regression (LR), Decision Trees (DT), Random Forests (RF), and deep neural network models were systematically assessed. Data preprocessing included feature selection based on importance ranking, Yeo–Johnson power transformation, and Principal Component Analysis (PCA) to improve model stability and performance. Among conventional ML models, LR combined with PCA achieved a coefficient of determination (R²) of 69.09% using MS data, while SVM yielded the best performance using RGB data (R² = 68.27%). Overall, deep neural networks outperformed other models, achieving R² values of 75.29% and 64.60% for MS and RGB data, respectively. Model performance varied notably across rice varieties; the Sona variety (n = 127) achieved the highest coefficient of determination (R² = 80.21% for MS and 76.34% for RGB), whereas varieties with fewer samples exhibited lower predictive performance. Results further indicate that ranking features by importance, rather than eliminating them, enhances predictive accuracy, particularly when using LR-derived feature importance, which proved critical for improving the performance of both LR and SVM models.
Accurate and timely crop yield estimation is fundamental for global food security, agricultural policy, and farm management. The Copernicus Sentinel-2 constellation has catalyzed a paradigm shift in Earth observation for agriculture, enabling field and sub-field scale monitoring. This review synthesizes recent advances in crop yield estimation that leverage Sentinel-2 data. A dominant theme is the transition from regional-scale to high-resolution field-level assessments, driven by three approaches: (i) empirical models using vegetation indices coupled with machine and deep learning (e.g., Random Forest, Convolutional Neural Networks); (ii) integration of process-based crop growth models (e.g., WOFOST, SAFY) through data assimilation of Sentinel-2 derived biophysical variables such as Leaf Area Index; and (iii) data fusion of Sentinel-2 with Sentinel-1 Synthetic Aperture Radar to overcome cloud cover. The synthesis shows that Sentinel-2-based frameworks can explain a large fraction of within-field yield variability, while performance remains constrained by limited ground-truth data, cloud gaps, and model transferability. Looking ahead, knowledge-guided models, self-supervised foundation-model pre-training, lightweight edge workflows, improved ground observations, and multi-sensor fusion are key pathways toward robust, operational decision-support tools for precision agriculture.
Background: Tropism, an adaptive growth mechanism often completely overlooked in tree phenotyping studies, is a crucial aspect of tree growth that allows them to reconfigure geometrically in relation to their immediate environment. This study introduces an integrated method to quantify tropic behaviour in plant phenotyping studies. Methods: The methodology combines cost-effective three-dimensional (3D) photogrammetric data capture from video, stem delineation techniques and 3D mathematical modelling of posture control for model-assisted identification of tropism traits. The proposed method was tested on a Pinus radiata D.Don. seedling subjected to a gravitational stimulus for 75 days. Stem posture was repeatedly measured using both 3D photogrammetry and fixed photography to create multitemporal 3D datasets and two-dimensional (2D) reference curves. Results: Individual 3D stem curves reconstructed with the proposed methodology introduced an error on spatial coordinates with a normalised RMSD ranging from 1.6 to 4.3% depending on time of capture, when compared with the 2D reference. The error for local tilt angle was higher than the error on spatial coordinates, with RMSD ranging between 5.6–12.2°, as expected for a first-order derivative. The gravitropic coefficient, capturing the sensing of and the reaction to local inclination by the plant, was underestimated by 2% if compared to the reference methodology. No contribution of autotropism (tendency to remain straight) was identified using the new methodology, but that contribution was found to be small using the 2D-approach and likely a key aspect of the gravitropic signature in the studied species. The major challenge with the proposed point cloud-based methodology arose from automated stem delineation. With dedicated algorithm enhancements to address stem occlusion in juvenile conifers and with more regular captures during plant motion, the proposed method could, however, perform identically to the 2D reference methodology. Overall, recovery of tropism traits performed equivalently whether using 2D or 3D data to fit the model of posture control. The minor discrepancies with experimental behaviour originated from fitting a simple kinematic model to complex real-world behaviour rather than data capture and digitising procedures. Conclusions: Overall, the proposed methodology, in its current form, offers a viable alternative to traditional 2D imagery methods at the cost of a small reduction in accuracy and capture time. The advantage of the 3D methodology is that it has the potential to track motion in multiple planes, whilst also measuring plant structure. With refinement, this methodology could be streamlined and adapted for deployment in field and operational environments at scale for phenotyping studies.
Reproduction assets foundThe paper's data availability statement explicitly deposits the raw photogrammetric point clouds and derived stem curves on Figshare and the R stem-extraction pipeline code on GitHub, both with public URLs.Dataset · publicthe
Ministry of Business Innovation & Employment (MBIE)
New Zealand as part of the Tree Interactions Programme
(Catalyst Fund C09X1923).
Supplementary materials and data availability
The raw photogrammetric point clouds and the stem
curves derived from both photogrammetry and 2D
imagery can be found at the following repository:
https://doi.org/10.6084/m9.figshare.32248617. The
R code for the stem extraction pipeline is available
at https://github.com/Robin-hartley/tropism-stem-curves-3d
Hartley et al. New Zealand Journal of Forestry Science (2026) 56:11 Page 14Open asset ↗figshare · 10.6084/m9.figshare.32248617pdf-raw-page:14 lines:97-113Code · publicamme
(Catalyst Fund C09X1923).
Supplementary materials and data availability
The raw photogrammetric point clouds and the stem
curves derived from both photogrammetry and 2D
imagery can be found at the following repository:
https://doi.org/10.6084/m9.figshare.32248617. The
R code for the stem extraction pipeline is available
at https://github.com/Robin-hartley/tropism-stem-curves-3d
Hartley et al. New Zealand Journal of Forestry Science (2026) 56:11 Page 14Open asset ↗github · Robin-hartley/tropism-stem-curves-3dpdf-raw-page:14 lines:97-113Plant phenotyping relevance matchCrossref · checked 15 Sept 2026
Aerial / UAVWhole plant / canopy / plot / fieldGrowth / time-series analysisTrackingVisualization / data managementGrowth / development / phenology
ABSTRACT Accurate monitoring of plant phenology is essential for agricultural decision‐making, as deciding the right time of fertilizer application, the maturity of the plant, and climate variation. Traditional manual monitoring often fails to capture the temporal variations across large fields. UAV‐based imaging combined with deep learning can provide a solution for automated phenological assessment. In this study, we propose an AI‐driven framework on UAV‐captured Indian mustard plants to predict the phenological progress. We applied deep learning methods, EfficientNet‐B0, and a proposed hybrid model of a Vision Transformer + LSTM to predict phenological growth from UVV‐captured images of the Indian mustard plant. Both models were trained under a supervised regression setup with extensive augmentation and optimization strategies. The results of the study show that the ViT + LSTM outperforms the EfficientNet‐B0 model in terms of prediction accuracy for mustard plant phenology. The R 2 = 0.9974, minimal errors MSE = 0.0111, and RMSE = 0.149 indicate that the ViT + LSTM provides more accurate phenological predictions on temporal and spatial dependencies. Correlation Analysis confirmed a strong linear and monotonic relationship with the ground truth, with r = 0.9989 and ρ = 0.9946. Gram‐CAM visualization showed that the ViT + LSTM captures the meaning area of the plant. These results were further validated using statistical measures such as Pearson's r and Spearman's ρ , which confirmed the reliability and consistency of the model's predictions.
Chao Mou · Jiahua Fan · Ang Liu · Chang Liu · Kecheng Huang · Huiyu Ding · Jingchen Li · Haiyan Zhang
Field / plotLiDAR / point cloudRGB / grayscaleRootStem / branchWhole plant / canopy / plot / fieldMorphology / geometry measurement2D/3D reconstructionVisualization / data managementArchitecture / morphology / geometry
Diameter at breast height (DBH) is a crucial indicator for obtaining tree phenotypes in orchard management, plantation monitoring, and agroforestry systems. LiDAR technology has high measurement accuracy, but it is costly and difficult to deploy flexibly in outdoor scenarios, while smartphones have emerged as a viable alternative due to their portability and low cost. In this paper, we propose a DBH estimation method based on monocular depth estimation, supported by a mobile application for algorithm deployment and result visualization. To address the limited computing resources on mobile devices, we design HR-DiffusionDepth, a lightweight diffusion-based monocular depth estimation network for smartphones, which generates pixel-wise 3D coordinates from a single image using camera intrinsics, thereby replacing LiDAR for DBH calculation. Experiments on the KITTI and SPREAD datasets show that HR-DiffusionDepth achieves the best depth estimation accuracy among similar lightweight models, reducing Abs Rel by up to 25.3% relative to the state-of-the-art (SoTA) lightweight baseline, with only 6.26 M parameters. The validation results show that the root mean square error (RMSE) of DBH estimation is 3.10 cm and the mean absolute error (MAE) is 2.25 cm, demonstrating the potential of this approach for agricultural scenarios such as orchards and plantations.
The increasing demand for agricultural production requires reliable and non-destructive methods for monitoring plant physiological conditions in real time, particularly in controlled environments. Spectral sensing in the visible to near-infrared (VIS–NIR) region offers a promising approach; however, the performance of low-cost sensors is often limited by calibration accuracy, wavelength-dependent sensitivity, and insufficient validation against plant physiological indicators. This study aims to evaluate the performance of a VIS–NIR plant monitoring system based on the AS7265x multispectral sensor in a controlled hydroponic environment. Lettuce (Lactuca sativa L.) was grown under two nutrient concentrations (600 ppm and 1200 ppm), and spectral reflectance data were collected across the 410–760 nm range. Sensor measurements were calibrated and validated against a LI-COR LI-180 reference spectrometer using linear regression, with performance assessed using the coefficient of determination (R²), root mean square error (RMSE), and spectral response consistency. The results show strong calibration performance, with wavelength-specific R² values ranging from 0.9682 to 0.9870 and RMSE values between 0.26965 and 5.19772. Although a systematic offset was observed, the AS7265x sensor preserved key spectral patterns, particularly in the green (510–560 nm) and red-edge (705–730 nm) regions. Differences in nutrient concentration were consistently reflected in both spectral responses and SPAD measurements, indicating sensitivity to plant physiological variations. These findings demonstrate that the AS7265x sensor provides reliable spectral information for relative plant monitoring and has strong potential as a cost-effective tool for plant quality assessment in controlled environments.
With the rapid development of artificial intelligence, UAV remote sensing, and agricultural Internet of Things technologies, crop growth monitoring is evolving from manual inspection and single-source analysis toward intelligent decision-making based on multisource perception. However, existing methods still suffer from limited robustness under environmental variations, insufficient integration between UAV imagery and sparse ground sensor observations, and weak capability for transforming predictions into practical agricultural management recommendations. This study proposes a UAV–ground sensor collaborative lightweight framework for crop growth assessment and agricultural decision support. The proposed framework integrates UAV RGB and multispectral imagery with ground sensor observations through a region-level aerial–ground alignment mechanism and a sensor-guided attention fusion module, enabling environmental conditions to enhance visual feature interpretation. Furthermore, a fact-constrained decision module is developed to generate management recommendations based on crop status, environmental risks, and field information. Experimental results demonstrate that the proposed method achieves superior performance in crop growth classification and yield-trend prediction, reaching Accuracy, Precision, Recall, and F1-score values of 92.47%, 91.86%, 91.39%, and 91.62%, respectively, with an RMSE of 0.381 and an R2 of 0.902. The lightweight framework requires only 6.18M parameters and 0.91G FLOPs, achieving 39.56 ms inference latency and 25.28 FPS on edge devices. The proposed framework also improves decision reliability, achieving an expert agreement rate of 89.34% and a risk identification accuracy of 90.18%. Economic analysis indicates that the proposed framework reduces labor cost, water consumption, and fertilizer input by 49.7%, 26.7%, and 23.0%, respectively, while increasing net benefit by 46.1% compared with conventional field management practices. These results demonstrate that the proposed method provides an accurate, interpretable, and deployable AI-driven solution for intelligent crop management in smallholder and medium-sized farming systems.
Introduction: Habitat restoration is necessary for the conservation and management of plant and animal species, especially in rare ecosystems. Drones may be well-suited to monitor changes in plant and animal communities in response to restoration efforts. The objective of the study was to examine whether drone imagery can detect differences in vegetation across multiple contexts. Materials and methods: Using a commercially available drone, I captured and processed aerial imagery with an open-source photogrammetric processing program. Point cloud data were processed to generate a vegetation density index, which was quantified across four cover types and compared between disturbance histories. In addition, using automated radio tracking, I compared vegetation density between used and available locations for Eastern Whip-poor-wills during the day and at night. Results: In August 2024, a drone flight covering a 3.05 km2 area of pine barrens captured 3372 images. Vegetation density differed by cover type (p = 0.001) and was greater in recently disturbed sites (p = 0.002). Scrub oak and recently burned sites had ~30% and ~12% greater vegetation density than deciduous forests and plots > 2 years post-disturbance, respectively. Vegetation density was lower at Eastern Whip-poor-will used locations than at available locations (151.0 vs. 159.7 points/m2, p < 0.001). Conclusions: Analysis of fine-scale differences in vegetation structure was important in discriminating subtle differences in habitat selection for Eastern Whip-poor-wills. This study demonstrated that drones and relatively simple image processing can be practical tools for restoration when quantifying and monitoring vegetation differences in dynamic ecosystems.
Reproduction assets foundThe paper's data availability statement points to a public Zenodo deposit containing the study's drone-derived vegetation density data and related measurements.Dataset · publicThe data supporting the findings of this publication has been made available within a publicly accessible
repository at https://doi.org/10.5281/zenodo.20398090.Open asset ↗Zenodo · 10.5281/zenodo.20398090pdf-page:11 lines:1-49Plant phenotyping relevance matchEurope PMC · checked 15 Sept 2026
Accurate monitoring of agronomic phenology is essential for yield estimation and food security assessment. However, currently available maize phenology datasets usually represent only a limited number of growth stages, restricting their application in process-based crop modeling and stage-specific agricultural management. Here, we present a high-resolution maize phenology dataset for Northeast China spanning 2001-2024 at 500-m spatial resolution and daily temporal resolution. By coupling MODIS spectral information with meteorological drivers in an energy-driven XGBoost framework, we retrieved eight key agronomic stages: Emergence, Three-leaf, Seven-leaf, Jointing, Flowering, Silking, Milking, and Maturity. Validation against observations from 91 agrometeorological stations during 2009-2024 demonstrates robust performance, with an overall RMSE of less than 5 days and R² values greater than 0.63 across all stages. Beyond overall accuracy, the dataset shows strong spatial consistency and temporal stability, preserves coherent regional phenological gradients, and captures interannual variations over the 24-year period. This long-term, multi-stage dataset provides a valuable benchmark for crop model calibration, climate change impact assessment, and the development of adaptive agricultural strategies in one of the world's major maize-producing regions.
Published30 Jul 2026The international archives of the photogrammetry, remote sensing and spatial information sciences/International archives of the photogrammetry, remote sensing and spatial information sciencesCited by 0 · OpenAlex ↗
Abstract. The harvest mouse, Micromys minutus (Pallas, 1771) is the smallest rodent in Japan and now listed in the Red Data Books of Tokyo, 2 prefectural capitals, and 28 prefectures in Japan due to drastic decline of grasslands. For the harvest mouse, the height and density of the tall grass species where nesting occurs are considered particularly important. However, it has been difficult to continuously and extensively acquire information on the three-dimensional structure of herbaceous vegetation. With recent development of UAV technology, UAV data are beginning to be applied to the analysis of herbaceous vegetation. For acquiring three-dimensional information via UAV, methods include using LiDAR sensors or generating 3D point cloud data from aerial photographs using SfM. This study evaluates whether UAV LiDAR or UAV SfM is more suitable for estimating the height of tall grass species such as Japanese silver grass (Miscanthus sinensis), which serve as important nesting sites for the harvest mouse. As a result of analysis, the proposed method was found to be effective to estimate grass height regardless of whether UAV LiDAR or UAV SfM is used. However, when comparing the accuracy of canopy height estimation using UAV LiDAR data alone, UAV SfM data alone, and combined UAV LiDAR and SfM data, combined UAV LiDAR and SfM data found to perform best. Maximum canopy height was found to be best estimated using the combination of median of hand-measured five maximum canopy height values and maximum height calculated using the combined UAV LiDAR and SfM data.
Abstract. Urban vegetation is essential for mitigating the Urban Heat Island effect, yet its cooling performance depends on its three-dimensional structure. This study combines high-resolution Unmanned Aerial Vehicle - based LiDAR (Zenmuse L2) and thermal imaging (Zenmuse H20) to analyze vegetation structure and surface temperature across 4 urban parks in San Nicolás de los Garza, Mexico. LiDAR data were processed to generate Digital Terrain Model, Digital Surface Model and Canopy Height Model models, enabling the segmentation of individual trees and extraction of structural metrics such as canopy height, crown area and point density. Thermal orthomosaics were co-registered with LiDAR models to quantify temperature contrasts between vegetated and impervious areas. Results reveal consistent cooling effects in all parks, with vegetated zones showing 8–15 °C lower surface temperatures depending on canopy density and maturity. Larger parks with continuous canopies displayed the strongest thermal regulation. This integrated LiDAR–thermal approach provides a precise and scalable framework for assessing microclimatic benefits of urban vegetation, supporting climate-resilient planning in rapidly urbanizing regions.
Introduction Timely and accurate detection of plant diseases is essential for ensuring global food security and supporting sustainable agriculture. Conventional diagnostic approaches, such as manual inspection and laboratory testing, are often time-consuming, labor-intensive, and impractical for large-scale or remote agricultural environments. Although deep learning models, particularly Convolutional Neural Networks (CNNs), have significantly improved automated plant disease classification, they often lack interpretability and struggle to generalize under diverse field conditions. Methods This study proposes AgriX-SENet, an explainable deep learning framework that integrates Squeeze-and-Excitation (SE) blocks with a DenseNet121 backbone to enhance disease classification performance. The SE blocks recalibrate channel-wise feature responses to emphasize disease-relevant information while suppressing background noise. To improve model transparency, Grad-CAM, SHAP, and LIME were incorporated to provide visual and feature-level explanations of the model’s predictions. The framework was trained and evaluated using the Plant Pathology 2020 dataset containing four classes: healthy, rust, scab, and multiple diseases. Results AgriX-SENet achieved a training accuracy of 97.47% and a validation accuracy of 95.07%, outperforming fourteen state-of-the-art deep learning models. The classification report demonstrated high precision and recall across most disease categories, although the scab class exhibited comparatively lower recall, indicating an opportunity for further improvement. The explainability analyses consistently showed that the model focused on pathologically relevant regions of leaf images, validating the reliability of its predictions. Discussion The proposed AgriX-SENet framework effectively combines high classification performance with model interpretability, addressing a key limitation of existing CNN-based plant disease detection systems. Its ability to provide accurate and explainable predictions makes it a promising solution for scalable agricultural diagnostics. Future work will focus on improving classification performance for challenging disease categories and optimizing the framework for deployment on mobile and edge computing devices to enable real-time field applications.
Reproduction assets foundThe paper trains and evaluates AgriX-SENet on the public Plant Pathology 2020 (FGVC7) Kaggle image dataset, which is the paper-specific plant image input for its disease-classification measurements. No author analysis code, trained model checkpoints, or supplementary code/data deposit is mentioned; the data statement (Dataset · publicPlant Pathology 2020 - Fgvc7 . Available online at: https://www.kaggle.com/competitions/plant-pathology-2020-fgvc7/data .Open asset ↗Kaggle · plant-pathology-2020-fgvc7lines:895-974Plant phenotyping relevance matchCrossref · OpenAlex · checked 5 Sept 2026
Accurate estimation of crop plant height using unmanned aerial vehicles (UAVs) is essential for field-scale crop monitoring and phenotyping. Most previous studies using UAV-based structure-from-motion (SfM) photogrammetry have relied on raster-based crop surface models (CSMs) and have evaluated their performance using accuracy metrics such as the coefficient of determination ( R 2 ) and root mean square error (RMSE). However, such evaluations provide limited insight into how estimation behavior varies across space and time, particularly during dynamic crop growth stages. To address this gap, this study conducted a time-series comparison of rice plant height estimates derived from UAV-SfM-generated dense point clouds (DPCs) and raster-based CSMs in farmer-managed paddy fields in Cambodia, which are characterized by heterogeneous micro-environmental conditions. Rice plant height was measured throughout the growing season and UAV-derived estimates were evaluated using regression analysis, analysis of covariance, and canopy cover dynamics. In the pooled analysis, both approaches achieved high overall accuracy, with R 2 = 0.92 and RMSE = 7.2 cm for the CSM-based approach and R 2 = 0.90 and RMSE = 8.8 cm for the DPC-based approach. However, time-series analyses revealed that CSM-derived plant height estimates exhibited strong location-dependent variability and sensitivity to early-stage canopy development, whereas DPC-based estimates showed more consistent performance across locations and growth stages. Regression coefficients derived from CSM-based estimates varied significantly among locations, whereas those from DPC-based estimates did not, suggesting that point-based representations may provide more spatially consistent estimation behavior under heterogeneous field conditions. By explicitly considering temporal dynamics, canopy development, and data representation, this study highlights the limitations of current raster-based UAV-SfM workflows for structurally complex crop canopies and suggests that DPC-based approaches may offer a useful complementary representation for crop monitoring and phenotyping, particularly when spatial consistency across heterogeneous field conditions is important.
The present study validates a custom multi-sensor system for high-resolution, non-phenotyping. The photogrammetric workflow was optimised by evaluating image density, algorithms, and camera calibration. Beyond validation, a case study demonstrated the system’s capacity to monitor early plant development following biostimulant treatment. Technical assessment revealed that prior internal camera calibration was unnecessary for the optics used. A reduced dataset of 120 images yielded reconstruction accuracies statistically comparable to full 360-image sets ( p > 0.05), confirming potential for maximised throughput efficiency by reducing processing time by approximately 66% without compromising data integrity. Analysis demonstrated that algorithmic settings determined reconstruction accuracy. Active parameter tweaks were essential for maximising surface area precision across all plant species ( R 2 ≥ 0.98). Conversely, volumetric accuracy required deactivation of these tweaks to maintain mesh consistency. While surface area estimation remained robust, volumetric precision showed species-specific variability driven by morphological complexity. Baseline parameters for accurate plant health scaling were established by defining species-specific vegetation index ranges (e.g., 0.38–0.82 for C. sativus ). The platform’s robustness was validated through a longitudinal study evaluating four treatments: yeast autolysate (A), a fungal biostimulant (F), their combination (AF), and a control (C). The system captured distinct morpho-physiological responses, demonstrating that autolysate-based treatments (A and AF) significantly enhanced biomass growth. The developed multi-sensor system recorded surface area expansions of 122% and 102% relative to the control ( p < 0.001), alongside a 110% increase in biological height. Fidelity of these 3D reconstructions was substantiated by a strong correlation ( R 2 = 0.96) between the 3D-derived leaf area index and ground-truth measurements. A key innovation of the pipeline is the integration of vertical distribution metrics as descriptive statistical tools, enabling high-resolution characterisation of canopy architecture. The plant health status metric evidenced enhanced physiological resilience in variants A and AF. Gravimetric analysis corroborated the non-destructive findings, confirming significant increases ( p < 0.001) in dried shoot weight of 77% (A) and 79% (AF). The validated system decoupled structural biomass from physiological health, offering broad utility across diverse phenotyping tasks. Such functionality streamlines the valorisation of industrial by-products into biopreparations, driving progress in sustainable agriculture.
Turfgrass phenotyping relies heavily on visual quality (VQ) ratings and RGB indices like DGCI, but these are limited by observer subjectivity, coarse ordinal scales, or ratio formulations that do not reflect perceptual color differences. Hyperspectral and machine-learning tools overcome some limitations but remain costly and difficult to generalize, motivating the need for scalable and interpretable RGB color metrics. We introduce ΔEg, a perceptually anchored CIELAB ΔE distance from an ideal green that provides a continuous and interpretable measure of canopy color evaluated alongside a panel of RGB-derived metrics. A 3 × 3 nitrogen × irrigation greenhouse experiment using hybrid bermudagrass (TifTuf, Cynodon dactylon × C. transvaalensis) quantified canopy responses with RGB imaging, spectral reflectance, CCM-300 fluorescence, and chlorophyll assays. ΔEg correlated strongly with chlorophyll (r = 0.72), similar to DGCI (r = 0.73), and both exceeded CCM-300 (r = 0.29). HSVi showed the strongest association with VQ (r = 0.84) and was most sensitive to irrigation (ηp2 = 0.63). CIELUV v* explained the greatest model variation (R2m = 0.94) and responded most to fertilizer (ηp2 = 0.84). The yellow fraction was significant across all main and interaction effects and captured canopy decline (r = −0.82 with VQ). An illustrative decision-support scenario using ΔEg indicated that moderate fertilizer combined with mild deficit irrigation optimized turf color and input efficiency. Conclusions apply to controlled conditions, with field-scale validation identified as future work. These results demonstrate that interpretable RGB color metrics, anchored by ΔEg, offer a scalable alternative to VQ scoring and spectral systems.
Reproduction assets foundThe paper's Data Availability Statement deposits the phenotype data and the authors' Python image-processing/metric-computation scripts and R statistical analysis scripts in the USDA National Agricultural Library Ag Data Commons, a public repository. The full RGB imagery archive, however, is only available upon requestCode · public2025;23:673–687. doi: 10.1002/lom3.10705.
Associated Data
Data Availability Statement
Data and Python scripts used for image processing and %G, %Gr, %Y, ΔEg, DGCI, HSVi, BA SD , CIELUV v* metric computation, and R scripts used for statistical analysis are be available in the USDA National Agricultural Library Ag Data Commons ( https://agdatacommons.nal.usda.gov/ ), Data for—Proxima Green: RGB Color Metrics for Turfgrass Phenotyping in Controlled Conditions, accessed on 27 July 2026. The full RGB imagery archive will be made available upon reasonable request.Open asset ↗USDA National Agricultural Library Ag Data Commonslines:691-695Plant phenotyping relevance matchOpenAlex · checked 5 Sept 2026
The ready-to-eat lettuce industry is rapidly expanding, increasing the need for reliable, scalable methods to assess seed germination and early growth under realistic soil conditions. This study presents an automated imaging-based approach for quantifying germination dynamics and seedling vigor using a low-cost multi-camera system under greenhouse conditions. Lettuce seeds were grown in soil either inoculated or non-inoculated with the soil-borne pathogen Rhizoctonia solani. Top-view images were acquired using commercial surveillance cameras and processed through a calibrated pipeline including geometric correction, color normalization, vegetation segmentation, clustering, and temporal tracking of emergence events. Seedling vigor was quantified through projected leaf area estimation. The proposed method enables accurate estimation of germination kinetics and growth dynamics under field-like conditions. Automated counts were validated against manual measurements at both intermediate and final time points, achieving high agreement in both cases. At the final assessment, the method reached R² = 0.98 and RMSE = 1.12, while at the midterm evaluation it achieved improved performance with R² = 0.998 and RMSE = 0.5, reflecting the lower complexity of plant structure at earlier growth stages. Results showed that pathogen inoculation significantly reduced both germination rate and seedling vigor, with up to 70% reduction in biomass accumulation. The proposed framework provides a robust, low-cost solution for high-throughput phenotyping of early plant development in soil-based systems, supporting scalable agricultural experimentation.
Abstract Deep learning has achieved remarkable success in rice disease diagnosis; however, existing methods often suffer from limited interpretability and poor robustness against open-world environmental noise. To address these challenges, this study proposes the Knowledge-Guided Multi-Task Rice Network (MTRNet) built upon a ResNet-50 backbone. Unlike conventional "black-box" models, MTRNet employs expert knowledge injection via a phytopathological matrix to explicitly disentangle disease features into Shape, Color, and Location attributes within a multi-head architecture. Furthermore, to mitigate false positives in complex field scenarios, a non-parametric Cascade Inference System (CIS)—comprising a biological grayscale filter and a visual consistency check—is introduced for robust Out-of-Distribution (OOD) detection and anomaly rejection. Experiments on a benchmark dataset of 5,932 field images, which primarily comprises four main rice diseases (Rice Leaf Blast, Brown Spot, Bacterial Leaf Blight, and Tungro), demonstrate that MTRNet achieves a diagnostic accuracy of 99.83%. Crucially, in an open-world robustness evaluation involving 1,000 non-agricultural noise samples, the proposed system achieved an 81.80% OOD rejection rate. By balancing diagnostic accuracy with structural transparency, this framework effectively narrows the gap between laboratory benchmarks and real-world agricultural applications.
Long-duration flood inundation can substantially suppress crop growth and cause yield loss, particularly in semi-arid agricultural regions increasingly affected by extreme rainfall. Timely crop damage assessment is critical for disaster response and insurance-related decision-making, but direct yield-loss observations are often unavailable during or shortly after flooding. This study proposes a phenology-guided regression framework for early crop damage assessment using multi-source SAR–optical observations. The study was conducted on the Tumochuan Plateau, Inner Mongolia, China, where severe rainfall beginning on 23 July 2025 caused widespread cropland inundation. Sentinel-2 EVI time series from 2022 to 2025 were fitted using a Savitzky–Golay (SG) filter, and annual area under the EVI curve (AUC) loss in 2025 relative to the 2022–2024 historical mean was used as a proxy for flood-induced crop damage. Optical features from Landsat-8/9 and Sentinel-2, together with SAR backscatter features from Sentinel-1, Lutan-1, and Gaofen-3, were incorporated into machine learning regression models. SAR features improved pixel-wise prediction, with the Random Forest model achieving the highest R2 of 0.62 using early-period features and 0.77 using later-period features. Village-scale aggregation further improved performance, yielding an early-period R2 of 0.84 across 123 and 0.78 across 122 villages. These results demonstrate the feasibility of SAR–optical and phenology-guided regression for early crop damage assessment under long-duration inundation.
Corn stunt is one of the most important diseases affecting maize (Zea mays L.) production in tropical regions of the Americas. The disease is caused by a complex of pathogens transmitted by the corn leafhopper (Dalbulus maidis), and its predominantly quantitative inheritance complicates the identification of tolerant genotypes under field conditions. In this context, we aimed to perform a comprehensive phenotypic stratification of corn stunt tolerance in a tropical public maize diversity panel and to identify contrasting inbred lines for breeding and genetic studies. A total of 360 inbred lines were evaluated under natural infection using three complementary disease-response traits: survivor plant health score (SPHS), proportion of survivor plants (PSP), and whole-plant health score (WPHS). Multi-trait mixed-model analyses revealed significant genotypic variation, moderate to high broad-sense heritability, and significant genotype × environment interactions for all evaluated traits. A multi-trait index (MSI), calculated from standardized best linear unbiased predictions (BLUPs), successfully integrated the three phenotypic components and enabled robust stratification of the diversity panel, identifying 60 highly tolerant and 60 highly susceptible inbred lines. Further, a genomic principal component analysis demonstrated that these phenotypic extremes were distributed across both tropical and subtropical germplasm, indicating that tolerance is not restricted to a single genetic background. The proposed phenotypic framework provides a robust and reproducible strategy for characterizing quantitative disease tolerance, identifying valuable parental germplasm, and establishing well-defined phenotypic extremes for future investigations of the genetic architecture of corn stunt tolerance.
Accurate and quick detection of plant leaf diseases is essential for precision agriculture to intervene promptly and boost crop yields. A new deep learning model called ResVNet has been introduced in this study. It combines the powerful local feature detection of ResNet152 with the global attention capabilities of Vision Transformer (ViT) and utilises Low-Rank Adaptation (LoRA) to accelerate fine-tuning. The PlantVillage dataset, which contains both healthy and diseased tomato samples, was used to train and test ResVNet. Experimental evaluation on the PlantVillage tomato dataset using stratified 5-fold cross-validation demonstrates that the proposed ResVNet model achieves a mean classification accuracy of 97.45%, along with superior macro-precision, macro-recall, and macro-F1 scores compared to existing deep learning architectures. The results of the confusion matrix and the ROC analysis validate its discriminatory power. The results highlight the potential of architectures strengthened with transformers in agricultural diagnostics. For real-time disease detection in the field, ResVNet is perfect for edge device deployment on drones and smartphones thanks to its high accuracy and adaptability. The application of Explainable AI (XAI) technologies for interpretability, integration with the Internet of Things (IoT), and multi-crop classification will all be explored in future studies. We will also look into model compression approaches so we can deploy efficiently in low-resource settings without sacrificing performance.
Abstract Pests and diseases are major constraints to cereal production, reducing crop yield, farm profitability, and food security worldwide. Timely detection of crop health threats and accurate assessment of infection severity are essential for effective crop protection, yet conventional field scouting remains labor-intensive, subjective, and unsuitable for real-time decision-making. Although recent advances in the Internet of Things (IoT) and deep learning have enhanced automated crop monitoring, most existing approaches focus on single-task disease classification and provide limited support for severity-aware management. This study proposes ResMDCL-PDM (Residual Network with Multi-Dimensional Compensation Layer for Pest and Disease Management), an IoT-enabled multi-task deep learning framework for precision pest and disease management in maize and rice production. The framework combines field-based environmental sensing with a modified ResNet-50 architecture enhanced by a Multi-Dimensional Compensation Layer (MDCL) to jointly identify crop species, classify pest and disease categories, and estimate infection severity. Field images collected from maize and rice farms at the Federal University of Agriculture, Abeokuta, Nigeria, were integrated with publicly available benchmark datasets. Following preprocessing and data augmentation, 8,556 annotated images were used for model development and evaluation. The proposed framework achieved an overall classification accuracy of 97.8% , outperforming AlexNet, VGG16, MobileNetV3, DenseNet121, EfficientNet-B0, and the baseline ResNet-50. High precision, recall, and F1-score, together with ablation analysis, confirmed the effectiveness of the proposed MDCL. The results demonstrate that integrating IoT-enabled monitoring with multi-task deep learning provides reliable, severity-aware decision support for targeted crop protection and offers a practical, scalable solution for sustainable precision agriculture.
Plant diseases have long been considered a major threat to global food production systems. Therefore, early diagnosis is vital to mitigate the risk of these diseases. This task can be challenging, as the number of harmful diseases is substantial. One technology that has gained widespread interest is artificial intelligence, specifically deep learning, which is used to identify plant diseases using leaf patterns. This paper presents two deep learning models, a custom CNN model and a transfer learning model based on the DenseNet-121 architecture. Experiments were carried out using the PlantCity dataset, which consists of twelve subsets of diverse crop species representing fruits, vegetables, and grains with variations among the subsets, including the number of classes, the subset sizes, class distribution, and visual complexity of disease symptoms. The two models were evaluated using multiple metrics, including accuracy, loss, precision, recall, and F1-score. Explainable AI using the LIME technique was deployed to better interpret the acquired results. Results showed that the developed transfer learning model based on DenseNet-121 had superior performance over the CNN model, with accuracies ranging from 86% to 99% across eleven experimented crops. In order to perform an independent experimental validation for the developed model, future work will focus on constructing a local crop dataset captured from Iraqi fields to evaluate the developed models based on the local environment.
Di Salvo JI, Joshi S, Hadadi M, Jubery T, Sarkar S, Singh A, Ganapathysubramanian B, Archontoulis S, Krishnamurthy A, Singh AK.
MilletNeRF / 3D Gaussian SplattingWhole plant / canopy / plot / fieldGrowth / time-series analysisVisualization / data managementGrowth / development / phenology
Finger millet ( Eleusine coracana (L.) Gaertn.) is a small seeded cereal known for its health benefits and its ability to grow in water-limited and saline environments. Despite its health and adaptation advantages, there is no standardized BBCH scale developed for finger millet, precluding standardization of a defined growth and development scale. The initial step in any breeding program and related research, is the precise, consistent and standardized description of crop growth stages. Therefore, we present a standardized phenological scale to describe and compare different stages of growth and development. In this study, we used four-finger millet accessions from the U.S.D.A. National Plant Germplasm System (NPGS) collection. Using a standardized BBCH chart as the baseline, we describe nine main stages of development to provide a simplified and user-friendly phenological growth staging scale for finger millet. A novel feature of this study was the use of Neural Radiance Fields (NeRF) to generate three-dimensional (3-D) renderings that align growth stages with 3-D imaging, paving the way for future phenological staging studies to incorporate 3-D visualization as a tool for standardization, researcher training, and reduction of observer bias across environments.
The Florida sugarcane industry is transitioning from manual to mechanical planting systems that use comparatively smaller seedcane pieces (billets) as planting material. A major limitation of mechanical planting is the increased seedcane requirement owing to mechanical damage and the increased vulnerability of seedcane pieces to soilborne pathogens that cause sett rots, particularly pineapple sett rot caused by Thielaviopsis spp. Current sugarcane breeding programs in Florida screen for major diseases, such as rusts, smut, ratoon stunting, and viruses, early in the breeding process but not for pineapple sett rot. This study aimed to isolate and identify Thielaviopsis spp. in the Everglades Agricultural Area (EAA), develop a single-bud inoculation protocol for greenhouse-based disease screening, and phenotype the current widely grown sugarcane varieties in Florida against Thielaviopsis spp. The pathogen was confirmed as T. ethacetica , consistent with previous reports from the EAA. A reproducible inoculation method was established and validated through symptom assessment, pathogen reisolation, and molecular confirmation. Using this protocol, six widely grown Florida sugarcane varieties showed significantly reduced germination (by more than 50%) and reduced above- and belowground morphological characteristics under infection, indicating susceptibility. Varietal differences were observed, with CP 03-1912 showing the highest mortality percentage and reduced growth under T. ethacetica infection. These findings highlight the vulnerability of current varieties to pineapple sett rot, especially under mechanical planting systems where smaller seedcane pieces are used. Furthermore, the developed inoculation protocol provides a scalable tool for early stage evaluation of resistance in breeding programs, offering potential to accelerate the development of varieties better adapted to mechanical planting.
Timely forecasting of maize productivity is essential to support precision agriculture and optimize management practices. In this study, we analyzed the potential of integrating ground-based measurements and UAV-derived spectral data for predicting maize grain yield (GY) under different fertilization conditions. Field data were collected at two key phenological stages: early vegetative stage (V7) and pre-harvest (R4). Ground-based measurements included SPAD, above-ground biomass (AGB), and leaf area index (LAI), while multispectral and hyperspectral imagery was acquired by drone. A series of Ordinary Least Squares (OLS) models was developed to evaluate the predictive performance of individual variables and their combinations. Model robustness was assessed using two validation strategies: Leave-One-Treatment-Out (LOTO) to assess model performance across the treatments included in the experimental design and random sampling to assess performance within the dataset. The results showed that yield prediction was less accurate during the early growth stages, where data fusion significantly improved the model’s accuracy (R2 = 0.82; MAE = 6.36 q ha−1; MAPE≈7 %). The predictive performance of VIs alone increased substantially in the pre-harvest stage, with the combination of red-edge indices and LAI proving to be the best model for late yield prediction (R2 = 0.86; MAE = 6.56 q ha−1; MAPE≈7%). Comparison of multispectral and hyperspectral data revealed comparable predictive performance, suggesting that multispectral sensors may already capture the key spectral information needed for yield forecasting. Furthermore, random validation consistently produced more optimistic results than the LOTO method, highlighting the importance of using validation strategies that explicitly account for the experimental design when evaluating model performance across the treatments included in the study. Overall, the present study demonstrates that yield prediction is highly dependent on the phenological stage and validation approach, and that integrating complementary data sources can improve model performance, particularly during the early growth stages. These findings should be interpreted as a proof-of-concept based on a single-site, single-season experiment with a limited sample size (n = 12), and therefore require further validation across multiple environments and growing seasons.
Background Covered smut in barley caused by Ustilago hordei leads to yield reduction and quality loss of stored grains and is especially challenging in organic production. However, screening for resistance remains challenging. The goal of our research was to evaluate protocols for screening covered smut in barley under normal and speed breeding conditions that could be scaled up for breeding purposes. We considered favorable pathogen growth conditions, a sufficient sample size to detect differences among genotypes through a power analysis, sources of disease escape or avoidance, and the infection effect on agronomic traits. Results In the first experiment, twenty genotypes treated with various inoculum concentrations were screened for disease incidence under a speed breeding system. Generally, low infection levels were found, likely due to disease escape or avoidance. Based on a power analysis, we modified the protocol to include more plants and improved pathogen growth conditions under a normal greenhouse system. With the modified protocol, the incidence of covered smut was significantly different among genotypes. The protocol also reduced the number of plants required to detect at least one infected plant. Artificial inoculation significantly decreased germination rates while head emergence, days to heading, and plant height were affected by disease infection in the most susceptible genotypes. We also found that covered smut incidence varied with tiller emergence order. The genotypes 'DH160779' (RES check), PI 270630', 'CIho15270', and 'MTV-color-158' presented potential resistance to covered smut. Conclusion The protocol has a high power to differentiate moderately resistant barley genotypes and we confirmed that specific agronomic traits were affected by disease incidence in susceptible genotypes.
Reproduction assets foundThe paper's disease-screening and agronomic-trait measurement data are publicly deposited on Zenodo, as stated in the Availability of data and materials section. No author analysis code or trained models are explicitly deposited.Dataset · publicThe data used and/or analyzed in the current study are available through the Zenodo, which is available at Gopinathan, G. (2025). Optimization of a protocol for covered smut in barley [Dataset]. Zenodo. [ 47 ] (https:/doi.org/ https://doi.org/10.5281/zenodo.17906264 ).Open asset ↗Zenodo · 10.5281/zenodo.17906264lines:190-223Plant phenotyping relevance matchCrossref · checked 14 Sept 2026
Qing Chang · Lixin Wang · Mallory L. Barnes · Darren L. Ficklin · Michael C. Benson · Kimberly A. Novick
Field / plotLeafWhole plant / canopy / plot / fieldObject detectionPhysiological trait estimationGrowth / time-series analysisLeaf traitsWater status / transpiration
Abstract Understanding the vulnerability of plants to more severe and frequent drought events and developing adaptive management strategies requires robust methods for quantifying long‐term changes in plant water stress (PWS). Most data‐driven explorations of long‐term trends in PWS have focused on alterations in canopy structure (e.g., leaf area index) or canopy structure‐dependent variables (e.g., gross primary productivity and evapotranspiration). This is largely because long‐term trends in canopy structure are relatively easy to detect from satellite observations. However, a focus on structural responses limits our ability to detect physiological stress due to challenges in isolating it from the effects of structural greening. Consequently, this difficulty hampers a comprehensive examination of long‐term PWS in the context of global greening trends. To address this gap, we developed a new process‐based metric for PWS to isolate physiological responses from structural greening, which we then used to detect global PWS trends over the past four decades. Combining site‐level and satellite observations at the half‐degree resolution across the globe, we found that accounting for greening‐related changes substantially alters the sign of long‐term PWS trends inferred from traditional approaches. Specifically, our study reveals a significant increase in PWS that is only detectable when accounting for structural greening trends. When greening trends are not accounted for, global PWS appears to have decreased over time. Overall, our results highlight the need to integrate structural dynamics and greening into PWS detection. Such an integration of observations and land models will improve our understanding of plant‐water‐energy interactions.
Abstract Detecting crop diseases early and responding promptly is vital for protecting agricultural productivity. It also helps maintain the quality and quantity of yields and reduces the risk of disease transmission to humans and livestock. Effective disease management is therefore critical to ensuring both global and local food security. However, traditional methods often based on visual inspection and delayed human judgment, are typically insufficient for identifying diseases at an early stage. Recent developments in Artificial Intelligence (AI) and the Internet of Things (IoT) offer new opportunities to address these challenges. By integrating IoT sensor networks with AI techniques such as machine learning and deep learning, it becomes possible to monitor plant health in real time and detect diseases with greater accuracy. This review explores the strengths and limitations of current AI-enabled IoT solutions in agriculture. It highlights how these systems leverage large-scale data and advanced image processing to outperform conventional methods in terms of speed, precision, and efficiency. Such improvements can significantly reduce crop losses and support more sustainable agricultural practices. Finally, the paper reviews key research trends, identifies current challenges, and outlines future directions in the field. It emphasizes the transformative potential of smart agriculture in advancing plant disease management and promoting environmentally responsible food production. This systematic review was conducted in strict accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines, synthesizing a final selection of 152 peer-reviewed papers. The overarching aim is to critically evaluate and map literature published between 2015 and 2026 using a systematic approach that addresses the integration of the IoT, AI, Machine Learning (ML), Deep Learning (DL), Convolutional Neural Networks (CNN), sensor technologies, and sustainable agricultural practices in the context of plant disease detection.
Alternanthera philoxeroides, an invasive alien species, spreads rapidly in river systems via vegetative propagation from stem fragments, requiring river-system-scale monitoring to understand its expansion dynamics and habitat preferences. This study used multi-temporal Sentinel-2 data to analyze spatio-temporal variations in fractional vegetation cover (FVC) within a 3.5 km river reach. FVC estimates derived from vegetation indices were validated against high-resolution aerial images, with an EVI-based model achieving the highest accuracy (RMSE = 9.2%), enabling reliable monitoring even in narrow (~24 m) channels. Time-series analysis from 2019 to 2024 revealed downstream expansion beginning in 2022. Annual maximum FVC (Cmax) was used to assess relationships with removal records and bank structures, showing that removal effects were temporary and more pronounced in the first year, while steel sheet-pile banks limited vegetation growth compared to concrete revetments. These results demonstrate that Sentinel-2 data can provide an effective and accessible tool for evaluating invasive plant dynamics and management effectiveness in low-flow river systems where A. philoxeroides dominates the floating vegetation community.
Yassine Ayat · Ali El Moussati · Oumayma Rachdi · Maryem Dinar · Abdelaziz El Aouni · Hajar Karkri · Mohammed Benzaouia · Wiame Benzekri · Ismail Mir · Aumeur El Amrani · Abdelmalek Mimouni
Field / plotWhole plant / canopy / plot / fieldStress / disease detectionPlant / canopy temperatureWater status / transpiration
Efficient irrigation management requires complementary information on atmospheric demand, soil conditions, and crop water stress. This study presents a low-power Internet of Things (IoT)-based irrigation system that integrates these components within a unified monitoring and control framework. The system combines LoRa communication, ESP32-based sensor nodes, soil and meteorological sensing, FAO-56 reference evapotranspiration (ET0), and canopy-temperature-based Crop Water Stress Index (CWSI). Irrigation decisions rely on the complementary use of ET0, in situ soil measurements, and CWSI rather than on a single indicator. A hybrid time-, event-, and query-driven acquisition strategy was implemented to adapt node activity and limit communication overhead. The system was deployed under outdoor conditions in Oujda, Morocco, demonstrating integrated sensing, wireless data transmission, crop-stress monitoring, and automated irrigation control. Energy characterization further showed distinct consumption profiles across sensing, communication, actuation, and low-power operating states, supporting the use of duty cycling to limit active node operation. The results demonstrate the feasibility of integrating environmental, soil, and crop-level information within a low-power IoT framework for adaptive irrigation management.
Introduction Climate extremes increasingly threaten agricultural production, yet many artificial intelligence systems in agriculture remain local, reactive and narrowly trained for one crop, region or sensing modality. Methods We present AgriFM, a multimodal geospatial foundation model that combines satellite image time series, radar, thermal observations, weather trajectories, soil properties, topography and sparse management variables to estimate crop-stress and yield-failure risk across crops and regions. AgriFM was pretrained using self-supervised objectives on 2.4 million field-season sequences and evaluated on a curated benchmark spanning maize, wheat, soybean, rice and sorghum across five agroclimatic regions. Results In held-out geography and time-split evaluations, AgriFM improved early stress detection and yield-failure prediction over statistical, crop-model and deep-learning baselines. The largest gains occurred during compound drought and heat events, for which AgriFM produced alerts 18 to 24 days earlier than the satellite-only baseline while maintaining improved calibration. Phenology-conditioned fusion improved transfer across planting calendars, and uncertainty calibration reduced false alerts at fixed recall. Discussion Because the study is based on retrospective datasets, these findings establish cross-region retrospective performance rather than prospective field efficacy. The results support further field-based evaluation of multimodal foundation models for climate-resilient crop monitoring.
The use of a combined assessment of the informational significance of vegetation indices for predicting the yield of spring wheat, taking into account varietal specificity and agrotechnical factors, has been studied. The test site was the field experience in the forest-steppe zone of the Novosibirsk Priobye. In the experiment, spring wheat of the Suenga and Novosibirsk 41 varieties was cultivated using intensive agricultural technology. For the analysis, data obtained using the DJI Phantom 4 Multispectral Phantom unmanned aerial vehicle during the crop growing period in 2023–2025 were used. Vegetation index values were calculated using five spectral channels: blue (B, 450 ± 16 nm), green (G, 560 ± 16), red (R, 650 ± 16), red edge (RE, 730 ± 16) and near-infrared (NIR, 840 ± 26 nm). For the analysis of informational importance, the following indices were used as predictors of crop yield: NDVI, NDWI, GNDVI, LAI, CVI, GCI, and ChlRE. For assessing the informativeness of the indices, independent methods were used: the F-statistic of one-way regression (ANOVA F-test), evaluation of mutual information (Mutual Information, MI), and feature importance of the random forest algorithm (Random Forest, RF). Each of the scores was normalized in the range [0; 1] using the min-max normalization method, after which a combined score was calculated as a weighted sum. For the Suenga variety, the stable predictors regardless of the experimental variants were CVI (tillering) and ChlRE (stem elongation and heading), while for Novosibirsk 41, the set of informative predictors significant ly depended on the combination of plant protection and fertilizer systems. It was found that chlorophyll content indices (GCI, ChlRE) increased the predictive relationship with yield under fertilization, while the water status index (NDWI) lost informativeness when fertilizers were applied.
Abstract Chili is an important economic and nutritional crop with a relatively limited availability of different disease-resistant varieties. Leaf diseases, including those caused by fungi, bacteria, viruses, pests, or nutritional deficiencies, significantly compromise production and crop quality. Early detection of these diseases is key to reducing yield loss; however, traditional visual examinations are limited by time constraints, human subjectivity, and low detection sensitivity at early stages of infection. To overcome these challenges, this study proposes a deep learning–based framework for early multi-class detection of chili leaf diseases using convolutional neural networks (CNNs). A real-field dataset comprising 16,392 high-resolution images of chili leaves across six disease and healthy classes was collected from multiple regions of Bangladesh. The dataset was preprocessed, augmented, and split into training and validation sets using an 80:20 ratio. Five pre-trained CNN architectures—DenseNet121, EfficientNetB3, MobileNetV2, ResNet50, and InceptionV3 were evaluated using a transfer learning strategy. Experimental results demonstrate that MobileNetV2 achieved the best performance, attaining an overall classification accuracy of 96%. The results indicate that the proposed system demonstrates strong generalization capability and effectively discriminates visually similar chili leaf diseases. This work can be considered a valuable application in precision agriculture, providing an efficient, automated, and practical approach for in situ early diagnosis of chili leaf diseases through smart farm management.
Above ground crop traits provide an early indication of a plant's capacity to tolerate stress, and are important for breeding programs aimed at improving stress tolerance. In this work, we present a high-throughput methodology to study morphological and physiological traits of individual quinoa plants over time under control, drought, and saline conditions. We used daily sideview imaging of individual plants, followed by segmentation of the panicle, leaf and stem using the deep learning U-Net++ segmentation model. The resulting segmentations were used in regression models to estimate leaf area, fresh and dry biomass, and leaf dry weight. The regression models showed high predictive accuracy. Using these estimates, we could calculate specific leaf area and leaf weight ratio. In addition, radiation use efficiency for above-ground biomass production was calculated, providing an independent physiological check on the consistency of these predictions. Finally, using automated measurements of plant transpiration we were able to determine daily averages of whole plant stomatal conductance. The results show that image-derived morphological traits can be used to accurately estimate biomass-related traits and to derive physiologically meaningful indicators of plant performance over time. This method provides a framework for non-destructive monitoring of quinoa responses to drought and salinity.
Precise characterization of alfalfa growth dynamics is essential for breeding accessions with superior regrowth capacity, persistence, and yield stability. However, traditional plot level and coarse scale observations suffer from low signal to noise ratios particularly before canopy closure when phenotypic data are strongly affected by weeds and soil background. Moreover, existing studies rarely capture the dynamic mechanisms of crop development across the entire growth cycle. To address this, the DINO-Pheno-Cluster framework is introduced as a foundation model driven and mechanism decomposed phenotyping approach. This decoupled framework first utilizes DINO-XMem, a few-shot individual plant segmentation network based on DINOv3 and a dual memory mechanism. It subsequently applies parameterized dynamic modeling guided by growth process knowledge. Validation utilized high frequency Unmanned Aerial Vehicle (UAV) imagery from 12 time points across three growing seasons covering 127 alfalfa accessions. DINO-XMem achieved an 89.54% mean Intersection over Union (mIoU) under a 10-shot setting and maintained 86.77% under extreme 1-shot conditions. It successfully resolved dense canopy oversegmentation outperforming fully supervised baselines by 4.08% to 11.97% in mIoU. Crucially, the extracted high purity time series trajectories were parameterized into specific biological indicators including maximum growth rate, comprehensive regeneration index, and seasonal stability index. Gaussian Mixture Model (GMM) clustering based on these mechanistic traits identified four distinct functional ideotypes comprising High yield/High regrowth, Upright/Sparse, High stability/Persistent, and Short/Dense, all validated by ground measured biomass. This workflow establishes a precision screening tool for multi harvest crops advancing crop phenomics toward process level analysis.
Blackberry micropropagation enables the rapid production of pathogen-free and genetically uniform plant material, although the evaluation of in vitro shoot development still relies on destructive and time-consuming measurements. This study investigated a low-cost smartphone-based 3D imaging approach for the non-destructive characterization of in vitro blackberry shoots (cultivar ‘Thornfree’) grown under different sucrose concentrations in the media (0, 7.5, 15, and 30 g L−1). Explants were cultured for 30 days under controlled environmental conditions in ventilated vessels containing 15 explants. Three-dimensional reconstructions generated using the viDoC RTK rover system coupled with an Apple iPhone 15 Pro Max were used to extract geometric traits, including shoot height, projected area, and shoot volume estimated through three complementary approaches, together with voxel-derived structural descriptors of shoot spatial organization and compactness. The proposed approach enabled the quantitative assessment of shoot architectural responses to sucrose availability, revealing differences in volumetric development and internal structural organization among treatments that would not be detectable by conventional measurements. The results highlight the potential of smartphone-based 3D phenotyping as a rapid, low-cost, and non-destructive tool for monitoring structural traits in micropropagated plant material and for supporting the optimization of in vitro culture conditions.
Plant phenomics has undergone rapid development over the past two decades, driven by advances in imaging, robotics, artificial intelligence and data analysis. Whilst field phenotyping is increasingly operational and scalable, the relevance of controlled-environment (CE) phenotyping is questioned because of concerns regarding the limited transferability of results to agricultural conditions. This Expert View first addresses the limitations and risks of using CE as surrogate of outdoor conditions. However, we argue that CE enables the disentangling of interacting environmental drivers allowing causal analysis of plant responses to multiple abiotic and biotic stresses. CE platforms also provide access to complex traits that are difficult or impossible to measure in the field whilst providing a robust framework in combination of field approaches to interpret and predict field performance. We further discuss contexts where CE remains indispensable, including quarantine and biosafety regulations together with emerging opportunities for agricultural innovation. Whilst limitations of CE systems are acknowledged, including issues of extrapolation, pot effects, environmental realism, and the indispensable need for rigorous envirotyping, we conclude that CE phenotyping should be regarded as an enabling analytical framework that complements and strengthens field phenomics for crop adaptation research under climate change.
Abstract The present work deals with Computer Vision precision in the agriculture domain designed to monitor the height variation of plants (h) and the percentage of Ground Cover (PGC). It brings together electronics and Computer Vision. the electronic part consists of two sensors: the first is the DHT11 sensor, which will monitor environmental parameters (temperature, humidity), and the second sensor is a camera (5 MP Raspberry Pi Camera Module Rev 1.3) which monitors the image acquisition to capture visual information about plants and Raspberry Pi 4 as the central processing unit for environmental data. For the Computer Vision part we have developed an algorithm able to do the acquisition and the segmentation of images acquired using Raspberry Pi 4 in real time.
Accurate estimation of crop water requirements is essential to improve irrigation efficiency for forage maize production. This study compared satellite- and UAV-derived normalized difference vegetation index (NDVI) models for estimating crop coefficients (K c ) and evaluated their operational performance for irrigation scheduling. K c -NDVI models were developed during the 2023 growing season and subsequently validated under field conditions during the 2024 season in two forage maize hybrids (N83N5 and Matador) under three irrigation strategies: conventional producer irrigation (ID1), satellite-based irrigation scheduling (ID2), and UAV-based irrigation scheduling (ID3). Both NDVI sources exhibited strong relationships with K c , with higher calibration accuracy for the UAV model (R 2 = 0.9414) than for the satellite model (R 2 = 0.8278). The UAV-based model applied 23-30% less irrigation water, maintaining high water productivity but also reducing crop growth, forage yield, and nutritional quality. In contrast, satellite-based irrigation scheduling promoted greater crop growth and produced the highest forage yield, reaching 59.8 t ha -1 in hybrid N83N5 while maintaining efficient water use. This treatment also improved forage quality by increasing dry matter and starch concentrations while reducing fiber fractions. The findings highlight the complementary potential of satellite and UAV imagery in precision irrigation and underscore the trade-offs between spatial detail, temporal resolution, and operational scalability. Furthermore, the results demonstrate that a stronger K c -NDVI relationship does not necessarily translate into improved irrigation scheduling performance. Under the conditions evaluated, the satellite-based model provided the best balance between water use, forage yield, and nutritional quality.
Abstract Architectural analysis provides a powerful analytical framework for understanding the ontogenetic trajectories of tree species and their adaptive responses to environmental constraints. Despite their ecological, economic, and cultural importance in West African agroforestry systems, the architectural development of Khaya senegalensis (Desr.) A. Juss. (Meliaceae) and Pterocarpus erinaceus Poir. (Fabaceae), two overexploited taxa classified as Vulnerable on the IUCN Red List, had never been formally described. This study presents the first complete characterization of their architectural development, from the seedling stage to senescence, based on architectural and retrospective analyses conducted on 360 individuals per species across seven localities along a south-north bioclimatic gradient in Côte d'Ivoire, covering contrasting vegetation zones ranging from dense humid forest to dry Sudanian savanna. Both species display well-defined ontogenetic trajectories comprising four phases: juvenile establishment, architectural construction, reproductive transition, and crown restructuring associated with ageing. Distinct architectural models were identified: K. senegalensis conforms to Rauh's model, characterized by a monopodial orthotropic trunk with indefinite growth and rhythmic acrotonic branching; P. erinaceus follows Troll's model, in which the orthotropic trunk progressively gives rise to a sympodial plagiotropic system with mixed terminal and lateral flowering. Architectural units, defined as the minimal structural organization enabling a species to reach reproductive maturity, were established at the adult stage: that of K. senegalensis comprises four axis categories and five branching orders, while that of P. erinaceus comprises three axis categories and up to six branching orders in old trees. Significant variation in growth-unit morphology among habitats and localities ( P ) revealed the architectural plasticity of both species in response to ecological gradients. The calculated Favourable Development indices ( FDi ) identified Bouaké and Katiola as optimal zones for K. senegalensis , and Bouaké and Toumodi for P. erinaceus , providing objective spatial criteria for reforestation planning. These findings demonstrate that architectural traits are robust indicators of development, adaptive strategies, and crown functioning. By linking structural organization to productivity, resilience, and regeneration potential, this study provides a scientific basis for integrating architectural analysis into reforestation programmes, sustainable forest management, and the design of agroforestry systems for threatened African tree species facing growing climatic and anthropogenic pressures. Complementary regression analyses further showed that phytomer number, rather than internode elongation, primarily governs growth-unit length in both species, and that growth unit diameter scales positively with growth unit length; a multivariate analysis of variance (MANOVA) confirmed that ontogenetic stage and locality, but not habitat alone, robustly structure growth-unit morphology.
Published23 Jul 2026The international archives of the photogrammetry, remote sensing and spatial information sciences/International archives of the photogrammetry, remote sensing and spatial information sciencesCited by 0 · OpenAlex ↗
Chaimaa Delasse · Dario Billi · I Gede Mahendra Darmawiguna · Kadek Ananta Satriadi · Arnadi Murtiyoso · H. Macher · Vincent Lecomte · Rafika Hajji · Tania Landes
Field / plotNeRF / 3D Gaussian SplattingThermalWhole plant / canopy / plot / field2D/3D reconstructionVisualization / data managementPlant / canopy temperature
Abstract. Urban trees provide critical ecosystem services in dense city environments, yet current workflows for monitoring their thermal behaviour remain confined to 2D desktop-based analysis with no three-dimensional spatial context or field-deployable visualization capability. This paper presents a complete pipeline for in-situ 3D thermal mesh visualization of urban trees in Augmented Reality (AR), combining Thermal InfraRed (TIR) image acquisition, Gaussian Splatting-based mesh reconstruction, quantitative validation, and mobile AR deployment. TIR images of a Tilia tomentosa acquired with a FLIR T560 camera are preprocessed with a standardized false-colour palette and fed into the MILo (Mesh-In-the-Loop Gaussian Splatting) framework to reconstruct a thermally attributed 3D mesh. Geometric evaluation against a Z+F IMAGER 5016 TLS reference using the M3C2 algorithm demonstrates that MILo recovers 13.5 times more canopy geometry than traditional multi-view stereo under thermal imagery, with a standard deviation of 4.0 cm. A colourmap inversion procedure recovers per-vertex temperature estimates from the GS-derived mesh colours, yielding a mean absolute difference of 0.7°C against direct T-Cam measurements (thermal camera mounted on the laser scanner), within the combined instrument accuracy of both sensors. The resulting thermal Gaussian Splat was deployed in a custom Android AR application supporting hybrid marker-based and GPS-based spatial anchoring for in-situ visualization. These results demonstrate the technical feasibility of GS-based thermal reconstruction and mobile AR as a medium for communicating three-dimensional canopy thermal information to educators and urban forestry practitioners.
Plant disease phenotyping underpins resistance breeding, epidemiology and crop-loss management, yet it remains a recognised bottleneck. This review asked whether the two metrics that dominate the discipline, the disease severity index (DSI) and the area under the disease progress curve (AUDPC), adequately represent disease as a temporally unfolding process, and what the evidence says about dynamic alternatives. Reporting followed PRISMA 2020 and the Synthesis Without Meta-analysis (SWiM) guideline. Web of Science Core Collection, Scopus, PubMed and a Google Scholar grey-literature sweep were searched for records published between January 2020 and December 2025, retrieving 1,192 records; 874 remained after de-duplication, 128 full texts were assessed and 31 studies met the eligibility criteria. Citation chasing added 24 foundational works, giving 55 included studies. Records were dual-screened (Cohen's kappa = 0.86), appraised with an adapted Mixed Methods Appraisal Tool, and synthesised using vote counting by direction of effect, an evidence map and structured cross-study comparison; meta-analysis was inappropriate because outcomes were not commensurable. Thirty studies (54.5%) represented disease at a single assessment and eight (14.5%) collapsed the epidemic into one integrated area, whereas only twelve (21.8%) retained the full trajectory. Across six outcome domains, all 29 study-level comparisons favoured the temporally richer method and none reported a null or negative result, an asymmetry indicating probable reporting bias. Certainty was high for visual-assessment findings, moderate for sensing and dynamic modelling, and low for field-realised genetic gain. The phenotyping bottleneck has migrated from data acquisition to data representation.
Published23 Jul 2026The international archives of the photogrammetry, remote sensing and spatial information sciences/International archives of the photogrammetry, remote sensing and spatial information sciencesCited by 0 · OpenAlex ↗
Abstract. Tree growth determines how much CO2 is sequestered from the atmosphere and temporarily stored in woody biomass. At the same time tree growth is affected by increasing temperatures, more frequent drought periods, late frosts and other extreme events associated with climate change. While continuous measurements of radial (secondary) tree growth using dendrometers are well established, monitoring of shoot elongation (primary growth) has largely been neglected because suitable measurement techniques are lacking. As a result, the effects of climate change on primary tree growth remain insufficiently understood. This work aims at reconstructing native deciduous trees in 3D as a basis for measuring and monitoring shoot elongation over entire tree canopies. Here we explored the use of low-cost UAV photogrammetry and of a multi-camera CraneCam system under real-world conditions. Data were collected in two study areas over an entire growing season. We present sensor evaluations, photogrammetric data acquisition and processing strategies. A special focus is placed on the analysis of the resulting photogrammetric 3D point clouds in terms of accuracy, resolution and completeness. Results demonstrate 3D point accuracies of 5-6 mm for entire trees using consumer-grade UAVs weighing less than 250 g and a 3D reconstruction completeness between 92% and 98% depending on the UAV type. The paper introduces a novel 3Dßprinted ground-truth branch to evaluate the capability to reconstructing fine-detail structures such as thin tree shoots. Finally, we discuss operational challenges and initial experiments towards a skeletonization of entire trees based on photogrammetric point clouds.
Published23 Jul 2026The international archives of the photogrammetry, remote sensing and spatial information sciences/International archives of the photogrammetry, remote sensing and spatial information sciencesCited by 0 · OpenAlex ↗
Abstract. Monitoring individual trees from unmanned aerial vehicle (UAV)-derived orthomosaics and digital surface models (DSMs) is important for forest management, but instance segmentation remains difficult in dense planted forests in Japan, where Structure from Motion (SfM)-derived DSMs often exhibit blurred crown boundaries, noise, and substantial variation in quality. Existing approaches can benefit from treetop information, but they are sensitive to threshold selection and may introduce false positives when incorrect treetop candidates are provided. To address this problem, we propose a method based on the polar coordinate transform and the fast Fourier transform (PFFT) that represents the local DSM shape around treetop candidates as compact descriptors and integrates them into Mask2Former. The descriptors are used both to suppress low-reliability candidates and to provide spatially meaningful treetop queries to the decoder, thereby improving the separation of adjacent crowns. We evaluated the method on Abies sachalinensis (Todo fir) plantation data acquired at two sites in Hokkaido, Japan, using Mask R-CNN and Mask2Former as baselines. Compared with standard Mask2Former, the proposed method improved mAP50 from 90.14% to 90.87%, mAP75 from 52.18% to 55.47%, and the F1 score at a confidence threshold of 0.5 from 89.86% to 92.08%. It also reduced the number of false positives by 41% without increasing false negatives. Qualitative results showed fewer over-merged crowns and better crown-boundary delineation. These results indicate that treetop-centered local shape cues are effective for instance segmentation in densely planted forests, although further validation across additional species and regions is needed.
- Enhancing agricultural productivity and attaining sustainable crop management depend on the early and precise identification of leaf disease. Using state-of-the-art technologies in precision agriculture like machine learning (ML) and image processing greatly increases the effectiveness of disease detection and facilitates well-informed decision-making. But conventional manual inspection techniques are still tedious, unpredictable, and prone to errors. In order to overcome these constraints, this research offers YOLOv12-CropNet, an innovative deep learning-based system for multi-crop leaf disease diagnosis in real time. The proposed YOLOv12-CropNet approach makes use of the Convolutional Block Attention Module (CBAM) for adaptive attention, the Content-Aware Reassembly of Features (CARAFE) up-sampling module to preserve fine-grained disease characteristics, the YOLOv12 architecture improved with Ghost Convolution for effective feature extraction, and Involution layers to capture spatially specific patterns. Inspection techniques are still laborious, arbitrary, and prone to mistakes. A substantial set of data of 38 classes of both healthy and sick leaves from a variety of crops, including tomato, potato, apple, grape, corn, mango and sugarcane, was put together for training and evaluation. Experimental results show that YOLOv12-CropNet finds a suitable balance between computational speed and accurate detection. Accuracy, F1-score, recall, and precision are important performance metrics that verify the model's resilience in challenging environmental and visual circumstances. The suggested technique provides a scalable and field-deployable way to assist effective identification of diseases and precision agricultural decision-making. The proposed YOLOv12-CropNet model exhibits better performance than the other evaluated models, attaining a 98.45% peak accuracy, 98.10% precision ,98.20 % sensitivity and a 98.18% F1 score, thereby highlighting its efficacy in multi-crop leaf disease detection.
Reproduction assets foundThe paper uses public Kaggle datasets as its phenotyping image inputs: the PlantVillage dataset (38 crop-disease classes) and the Sugarcane Leaf Disease dataset, both cited with explicit public URLs. No author code, models, or checkpoints are reported as publicly available.Dataset · public[37] PlantVillage Dataset. Available online: https://www.kaggle.com/datasets/abdallahalidev/plantvillage-datasetOpen asset ↗pdf-page:22 lines:1-61Dataset · public[39] Sugarcane leaf Disease Dataset available online: https://www.kaggle.com/datasets/nirmalsankalana/sugarcane-Open asset ↗pdf-page:22 lines:1-61Plant phenotyping relevance matchEurope PMC · Crossref · checked 5 Sept 2026
Abstract The diseases on sesame leaves have a huge implication on the production and earnings of farmers particularly in the developing areas. It is important to ensure that the disease is properly managed by identifying it early and correctly. The purpose of this study is to create a deep learning-based framework that could be used to categorize four of the most common scenarios involving sesame leaves, i.e. Healthy Leaf, Leaf Spot Disease, Yellowing Leaf Syndrome, and Leaf Damage by Insects using high-resolution images acquired in Pabna, Bangladesh. The preprocessing, augmentation and split of a set of 3,540 images were performed into training and validation sets. The pre-trained convolutional neural networks models were trained and tested on five models inceptionV3, EfficientNetB3, ResNet50, MobileNet, and DenseNet121 by measuring the metrics such as accuracy, precision, recall, and F1-score. MobileNet achieved the highest accuracy of 96.33%, precision of 96%, recall of 96%, and the F1-score of 96%, which is the best amongst them. The findings indicate that deep learning architectures are capable of classifying the sesame leaf diseases in a reliable and precision-oriented way that is not affected by different environmental circumstances. The study facilitates the creation of the automated, efficient methods of detecting the disease at an early stage, cutting down the number of pesticides used and enhancing crop control. Future direction will be to enlarge the dataset, add temporal data and to implement lightweight models so that it can be deployed to real-time field projects
Precision agriculture increasingly relies on detailed structural information, such as canopy height and canopy volume, to enhance crop health monitoring and operational safety. However, existing methods based on costly LiDAR or RGB-D sensors are often impractical for large-scale deployment in dynamic and unstructured horticultural environments. Furthermore, conventional 2D segmentation and SLAM-based pipelines typically generate sparse, geometrically inconsistent semantic maps which are insufficient for actionable structural analysis in agricultural applications. To overcome these limitations, we propose a monocular 3D structural mapping framework tailored for horticultural plants via semantic scene completion. At inference, the proposed model takes a single RGB image as input and predicts voxel-wise geometry and semantics, from which task-oriented structural maps, including canopy height, canopy volume, and obstacle-aware traversability layers, are derived. Specifically, we first introduce a Depth-Aware Decoder Module that explicitly recovers depth in the spatial domain and fuses 2D-to-3D features, thereby mitigating depth ambiguity and reducing reliance on accurate pose. Second, an NCS-Guided Geometry Encoder is designed to inject normalized depth into voxel positional embeddings, enabling self-attention to perform global relational modeling within a depth-aware geometric coordinate system. In addition, a Global Encoder is utilized to refine local structural details, while an occupancy head produces the final 3D semantic completion outputs. We construct a horticultural 3D semantic scene dataset using an RGB-D sensor, which serves as a benchmark for evaluating our method, while the deployed model remains RGB-only. Extensive quantitative and qualitative experiments are conducted on both the Semantic-KITTI dataset and our dataset. On our dataset, the method achieves 82.31% occupancy IoU, 84.26% mIoU, and 86.25% precision. Beyond voxel-level evaluation, manual field measurements further show canopy height MAE values of 0.019-0.026 m and canopy volume proxy relative errors of 8.4%-11.4%. These results demonstrate the effectiveness of our approach in real-world agricultural scenarios, providing actionable structural insights for crop monitoring and autonomous robotic operations.
Introduction Safeguarding the yield and quality of field crops against abiotic stresses is critical for large-scale agricultural production and precision agronomy. This study addresses the challenges of detecting concealed fruit stress and overcoming label scarcity in unmanned aerial vehicle (UAV) multispectral monitoring, using blossom-end rot (BER) in processing tomatoes as a case study. Methods We developed a leaf-fruit synergistic Roll and Color Disease Index (RCDI), integrating fruit incidence with visible canopy phenotypic features to characterize the combined canopy-fruit stress status associated with BER. A three-level screening strategy was used to identify the optimal spectral feature set. A Multi-model Collaborative Cyclic Self-Training (MCC-ST) framework was subsequently developed to address the limited availability of severity-labeled samples. Results The combination of GRVI, NDVI, and SAVI was identified as the optimal spectral feature set, achieving stable within-dataset binary classification performance of approximately 97% in repeated cross-validation. Under 30 random-seed repeated stratified three-fold cross-validations, MCC-ST + DT and MCC-ST + RF achieved RCDI-based BER severity-grading accuracies of 85.00% ± 0.31% and 85.07% ± 0.42%, respectively. Compared with the corresponding original DT and RF models, MCC-ST improved repeated-validation accuracy by 16.19-4.09 percentage points. Discussion The RCDI helps bridge the observational gap between canopy signals and concealed fruit stress, while MCC-ST alleviates the bottleneck associated with label scarcity. The proposed approach provides a promising framework for crop abiotic-stress monitoring under limited-label conditions.
Accurate color representation is critical for UAV-based crop phenotyping, yet UAV images are often distorted by variable illumination and camera exposure settings. Here, we propose Lite U-net FiLM (LUF-net), a lightweight U-net framework integrated with feature-wise linear modulation (FiLM) layers, which incorporates both multispectral-derived irradiance and camera exposure parameters as auxiliary inputs. This metadata-aware design enables dynamic modulation of intermediate image features, allowing the network to disentangle unified canopy color from environmental artifacts. Experiments conducted across diverse crop types (soybean, rice, maize), flight altitudes (6m, 12m, 25m, 40m), and illumination conditions (overcast skies, cloudy, sunny, early morning) demonstrate that the LUF-net model substantially reduces the mean absolute percentage error to below 5.5%, outperforming conventional Gray-world, U-net, AlexNet-MLP, and SIDBlock-MLP methods. Ablation experiments further show that both irradiance and exposure metadata provide complementary information, while FiLM-based conditioning effectively integrates these acquisition parameters into feature learning, jointly contributing to improved color reconstruction performance. Moreover, correlation analysis indicates that the model's color reconstruction errors are weakly dependent on irradiance and exposure settings, confirming that LUF-net reduces sensitivity to external imaging conditions while maintaining physically meaningful and robust corrections.
Reproduction assets foundThe paper's data and code availability statement explicitly points to a public GitHub repository containing training/evaluation/inference code, pretrained model weights, and example data for the LUF-net color calibration method.Code · publicThe data and source code for model training, evaluation, and inference, together with pretrained model weights, example data, and detailed usage instructions, is publicly available at: https://github.com/wangchufeng3652/color-correction.Open asset ↗wangchufeng3652/color-correctionhtml-lines:278-299Plant phenotyping relevance matchEurope PMC · Crossref · checked 5 Sept 2026
Abstract Outbreaks of plant diseases are major threats to world food security particularly in areas where real-time monitoring and quick decision support are constrained by low-power edge gadgets and untrustworthy connectivity. In order to overcome these issues, this paper presents a FPGA-Accelerated IoT implementation of a Causal-Attention Multi-Modal Deep Learning Network, named EpiFusionNet-Edge, that can be applied to monitor crop diseases with real-world farming scenarios with high precision and scalability. The framework incorporates five data modalities that are complementary in nature and they include RGB leaf pictures, microscopic foldscope images, UAV hyperspectral signatures, microclimate IoT sensor measurements and region-specific pathogen/pest pressure indexes giving a complete picture of the health of the plant. Dual causal-attention mechanism is proposed to simulate both spatial and temporal environmental factor activation, which helps to detect and make predictions at the early stage and provides an explanatory logic behind the decisions. Multi-task learning enables classification of diseases, quantification of their intensity at the level of a micro-prediction and prediction of outbreaks in the short term (1–30 days). In order to achieve deployability in resource-constrained settings, the proposed deep learning architecture is ensemble-distilled, structurally pruned, and INT8-quantized, and hardened on a Xilinx Zynq-7000 FPGA platform. The FPGA accelerator is 43.2x faster inference, 88 percent less power usage, and less than 10 ms latency, which allows real-time execution of continuous field monitoring with IoT sensors. Cross-condition assessment on multi-domain datasets shows that there are great improvements on cross-environment generalization rates with 98.6% classification accuracy, 92.7% severity estimation accuracy and less than 3.5% degradation with domain shift. Grad-CAM + + and causal feature traceability further add interpretability with the focus of the model and the pathological indicators proven by experts. The findings show the promise of using a combination of IoT sensing, multi-modal AI fusion, and FPGA hardware acceleration to develop a deployable and scalable and transparent system with regard to precision agriculture. This paper creates a roadmap to a new generation of smart farming systems that are able to conduct disease surveillance and actively protect crops at the periphery in an autonomous manner.
Abstract Quantifying blueberry fruit yield and maturity is important for evaluating yield potential in breeding trials, but manual measurement remains slow, labor-intensive, and costly. Object detection and classification networks offer a high-throughput solution, yet few studies validate image-based counts against hand-harvested ground truth while explicitly accounting for canopy occlusion. Hence, this study developed a multi-class berry detection pipeline for immature and mature berries and validated image-based estimates against hand-harvest counts across 32 diverse southern highbush blueberry genotypes. Among the models evaluated, YOLOv8x achieved the highest detection performance, with an mAP50 of 0.82 and an mAP50–95 of 0.66. External validation produced F1 scores ranging from 0.74 to 0.91 for berry maturity classes. However, image-based detections systematically underestimated hand-harvested fruit counts, with R² values ranging from 0.40 to 0.61. Fruit occlusion varied widely among genotypes, from 42% to 90%, indicating that canopy structure strongly affects berry visibility. Incorporating image-derived canopy architecture, color, and texture features improved predictions of berry counts and maturity. Partial Least Squares regression provided the best performance, increasing R² values of hand-harvested fruit counts, ranging from 0.52 to 0.74. These results show that accounting for canopy occlusion improves image-based estimation of blueberry yield and supports more accurate high-throughput phenotyping.
Agrophotovoltaic (APV) systems provide a unique opportunity for improving agricultural land-use efficiency by combining crop production with solar energy capture via photovoltaic panels. In-depth information on plant growth patterns within the spatially heterogenous microclimate created by APVs would enable better planning and management within such unconventional systems. Thus, the present study demonstrates the implementation of a customized robot-mounted 3D-multispectral imaging system for monitoring the growth and spectral reflectance patterns of a conventional soybean cultivar "Eiko" (EK) and a chlorophyll-deficient mutant variety MinnGold (MG) under an APV system. Weekly trends in canopy morphometric features revealed significant variations in canopy height, surface area, light penetration, and volume across the APV field depending on the proximity with the overhead solar panels for both EK and MG, with plants receiving adequate rainfall and intermittent shade performing the best. Furthermore, although spectral indices exhibited variations between EK and MG due to intrinsic differences in pigmentation, symptoms of stress could be detected for both genotypes within rain-shaded areas of the APV plot. Hence, the present investigation depicts the potential for complementary usage of robotics and machine vision for high-precision high-throughput crop monitoring under APVs, which would help improve crop management within such non-homogenous cultivation systems.
Introduction Accurate and non-destructive estimation of wheat biomass is essential for crop growth monitoring, yield prediction, and precision agriculture. Unmanned aerial vehicle (UAV)-based remote sensing, integrating both spectral and structural information, has shown great potential for biomass estimation. However, the mechanisms by which different types of variables contribute to biomass prediction remain poorly understood, especially when using machine learning models. Methods In this study, we fused spectral reflectance, vegetation indices, and canopy height data derived from a UAV multispectral camera to estimate wheat biomass across four growth stages (jointing, booting, heading, and filling). Four machine learning algorithms-XGBoost, Random Forest Regressor (RFR), Support Vector Regressor (SVR), and LASSO-were employed and compared. Results and discussion The results showed that XGBoost achieved the highest accuracy (R 2 = 0.919, RMSE = 102.43 g/m², MAE = 77.43 g/m², RRMSE = 19.71%). Furthermore, SHAP (SHapley Additive exPlanations) analysis revealed that canopy height (CH) was the most important variable, followed by spectral indices such as R842 and GNDVI. The univariate and global contribution analyses demonstrated that structural and spectral variables played complementary roles in biomass estimation. This study provides a mechanistic understanding of variable contributions and offers a robust framework for UAV-based wheat biomass estimation.
Abstract Blanket-rate agrochemical scheduling — a practice wherein the same quantity of fertilizer or pesticide is spread uniformly across an entire field irrespective of spatial or temporal crop need — persists as the dominant farm management paradigm across rural India and large parts of South Asia. This approach generates cascading inefficiencies: excess nitrogen drains into waterways, off-target pesticide deposits devastate pollinators, input costs erode thin profit margins, and wide-scale greenhouse gas release from soil microbial activity accelerates climate change. The study documented here addresses this challenge through a purpose-built, four-layer intelligent field management platform. The platform ingests continuous data from drone-mounted multispectral cameras, in-field IoT soil probes, a wireless weather station, and cloud-sourced Sentinel-2 satellite imagery, then passes these inputs through a cascaded AI inference stack. A fine-tuned YOLOv8-L network performs real-time pest and foliar disease localisation; a ResNet-50 backbone quantifies canopy health across five stress gradients; a two-layer stacked LSTM projects short-horizon yield trajectories; and a Deep Q-Network autonomously plans drone spray routes weighted by field-specific prescription maps. Field validation spanned two consecutive growing seasons (Rabi 2022–23 and Kharif 2023–24) across six georeferenced plots covering 4.8 ha at Baramati, Maharashtra. Outcome metrics recorded during head-to-head comparison with conventional practice included a disease detection score of 95.6% mAP, a 47.3% reduction in total nitrogen applied, a 38.1% decrease in pesticide volume, and a 22.4% uplift in harvested grain weight. Together, these field-verified numbers substantiate the operational readiness of integrated AI precision agriculture for smallholder deployment.
Reproduction assets found保存済みの本文根拠を更新済みルールで再検証し、公開資産1件を確認しました。Code · publicData Availability The annotated image dataset (14,300 images, 23 classes), trained YOLOv8-L and ResNet-50 weights, LSTM model files, DQN policy checkpoint, and all analysis scripts are archived at https://github.com/precision-agri-ai (Zenodo DOI: 10.5281/zenodo.XXXXXXX).Open asset ↗precision-agri-ailines:161-182Plant phenotyping relevance matchCrossref · checked 5 Sept 2026
Remote sensing has become an important tool for crop monitoring and precision agriculture, yet its applications in sugar beet production remain fragmented across sensing platforms, target traits and modelling strategies. This review synthesises the development, current applications and future directions of remote sensing in sugar beet production, with particular attention to the transition from crop monitoring to precision management. A structured search was conducted in Scopus and the Web of Science Core Collection for publications from 2003 to 2025, and 181 relevant peer-reviewed articles were retained for thematic analysis. The literature shows a clear increase in sugar beet remote sensing studies, particularly after 2015, coinciding with the availability of Sentinel-2 imagery and, from 2016 onwards, the growing use of unmanned aerial vehicle-based sensing. It also indicates a gradual shift from crop mapping and canopy monitoring towards disease detection, weed mapping, yield prediction and management-oriented applications. Current studies demonstrate the value of satellite, unmanned aerial vehicle and proximal sensing for retrieving canopy traits, assessing biotic stresses, estimating root yield and supporting field-scale management. However, sugar beet presents specific challenges because its economic value depends not only on canopy development or root biomass, but also on sucrose concentration, recoverable sugar yield, and processing quality. These quality-related traits remain less studied and are difficult to infer directly from canopy observations. Modelling approaches have evolved from vegetation-index-based empirical models towards machine learning, deep learning, multi-temporal analysis, data fusion and crop model assimilation, but issues of model transferability, ground-truth availability and operational decision support remain unresolved. Future research should strengthen multi-source observations, external validation, quality-oriented prediction and decision-support workflows to promote robust, scalable and economically meaningful remote sensing applications in sugar beet production.
Smartphone-based precision spraying aims to restrict chemical application to target organs, but it also requires an organ-level perception model that operates on RGB alone at deployment, without dedicated depth sensors. Combining with monocular depth foundation models such as Depth Anything V2, RGB images can be enriched with dense pseudo-depth during training, turning close-range crop segmentation into a privileged-depth-guided learning problem where depth is available at training but absent at inference. Based on such close-range handheld capture, intra-foreground depth separability collapses while only the foreground–background depth gap remains partially preserved, exposing an imaging regime that we refer to as Close-Range Depth Degeneracy (CRDD) and that undermines the discriminative value of monocular depth for organ segmentation. Even with the available privileged-depth signal, existing RGB-D fusion and depth-branch hallucination routes still either introduce inference-time depth dependence or fail to retain the depth-conditioned effect inside RGB-side parameters under CRDD. How monocular pseudo-depth can serve as training-only privileged information for RGB-only close-range crop spraying under CRDD remains insufficiently studied. To this end, this paper proposes a privileged-depth modulation framework, MDM-Seg, by converting monocular depth from an inference-time modality into a CRDD-conditioned training signal to remove the inference-time depth dependence. Specifically, we characterize CRDD and propose a four-indicator depth-usability diagnostic that delivers a pre-training verdict on the target dataset, with verdicts validated against post-training segmentation outcomes. Then, we formalize the CRDD-conditioned training intervention as a depth-conditioned affine network (DepAN) that injects depth-gradient cues into decoder features, and define a depth hard-pixel mining loss (DHPM) to reweight per-pixel cross-entropy with depth-value evidence, both operating inside the main segmentation stream. Finally, we deploy MDM-Seg on RGB input alone and derive a lightweight consistency trace between the RGB-only and depth-conditioned heads, producing an image-level reliability cue for targeted-versus-fallback spraying without depth at inference. Extensive experiments on Pepper-Field-3940 demonstrate that our MDM-Seg achieves effective depth-conditioned RGB-only segmentation under CRDD, i.e., depth-conditioned modulation retained inside RGB-side parameters with no inference-time depth dependence, such as mIoU 0.8995 with the ResNet-101 configuration ( + 3.73 pp over the RGB-only DeepLabV3+ baseline), a lightweight training-side DepAN module, and an image-level reliability cue rank-correlated with segmentation quality (Spearman ρ = 0.74 ). On a Jetson Orin NX 16 GB with TensorRT FP16, the complete deployment-oriented MobileNetV2 dual-head prescription pipeline runs at 23.87 ms per image, corresponding to 41.89 FPS, 0.62 GB peak memory, and 0.56 J per image.
Accurate identification of individual plants from unmanned aerial vehicle (UAV) imagery is essential for high-throughput phenotyping and data-driven decision-making in plant breeding. This study presents MatchPlant, a modular, open-source Python pipeline with a graphical user interface for UAV-based single-plant detection and geospatial trait extraction. The pipeline integrates UAV image processing, user-guided annotation of selected undistorted images, convolutional neural network–based object-detection training, forward projection of bounding boxes onto an orthomosaic, and shapefile generation for spatial phenotypic analysis. This workflow preserves native image geometry during detection while maintaining coordinate traceability from source imagery to georeferenced outputs. Across five independent training runs using early-season maize imagery, MatchPlant achieved source-image-level detection performance of AP@0.5 = 90.3 ± 1.1% and mAP@0.5:0.95 = 43.4 ± 2.9%. Orthomosaic-level evaluation after forward projection showed AP@0.5 = 89.7 ± 0.8% and recall = 93.7 ± 1.2%, demonstrating the workflow’s ability to transfer plant detections into georeferenced outputs. Plant-level traits, including plant height derived from canopy height models and NDVI derived from vegetation index rasters, showed strong agreement with manual annotations ( r = 0.87–0.97). Detection outputs were reused across time points with minimal additional annotation, supporting temporal phenotyping during early growth. The framework was validated using maize imagery from a single site and growing season, where plant separation remained clear. By combining modular design, reproducibility, and coordinate traceability, MatchPlant provides an open-source workflow for UAV-based plant-level analysis, with broader applications requiring validation across additional crops, sensors, growth stages, GSDs, and field conditions.
Reproduction assets foundThe paper's MatchPlant analysis pipeline is publicly available on GitHub, and the maize case-study training dataset and pre-trained model are publicly available on Zenodo; both are paper-specific, public, and actionable.Dataset · publicThe public datasets supporting the case study are available on Zenodo at https://doi.org/10.5281/zenodo.14856123 (accessed on February 14, 2025).Open asset ↗Zenodo · 10.5281/zenodo.14856123lines:169-250Model / weights · publicThe training dataset and pre-trained model used in the maize case study presented in Section 3 are also publicly available via Zenodo ( Sangjan et al., 2025a ) at https://doi.org/10.5281/zenodo.14856123 (accessed on February 14, 2025).Open asset ↗Zenodo · 10.5281/zenodo.14856123lines:70-82Code / dataset availability confirmedEurope PMC · checked 5 Sept 2026
This study compiles and releases the first standardized rapeseed phenology observation dataset spanning forty-four years (1981-2024) over the core winter rapeseed production region of the Middle and Lower Yangtze River Plain in China. The data originate from systematic observations at 50 national-level agrometeorological stations across six provinces: Jiangsu, Zhejiang, Anhui, Jiangxi, Hubei, and Hunan. The dataset provides complete records of the specific dates for each phenology stage from sowing to maturity, including eight key phenology periods: Sowing (SO), Emergence (EM), Five-leaf (FV), Bud Formation (BF), Stem Elongation (SE), Flowering (FL), Green Ripening (GR), and Maturity (MA), along with the calculated durations of six distinct growth lengths. We implemented a multi-level quality control protocol encompassing internal logical checks, statistical outlier detection, climatological validation, time series homogenization, and expert arbitration. This protocol effectively constrained data uncertainty and corrected non-climatic discontinuities. Univariate linear regression was further employed to quantify the decadal change trends of each phenology period and growth length, supplemented by Kernel Density Estimation (KDE) to characterize their probability distribution features. The final dataset is presented as structured tables (in xlsx format) and high-resolution diagnostic plots (including trend and density plots), with a total volume of approximately 470 MB, systematically organized by province and station. This dataset fills a critical gap in long-term, standardized rapeseed phenology data for the region. The integrated analysis of phenology dates, growth stage durations, and their trends across the entire network provides an indispensable, high-quality empirical foundation. It is designed to support in-depth investigations into the nonlinear response mechanisms of overwintering crops to climate warming, improve crop model parameterization and validation, and inform regional adaptive management strategies.
Reproduction assets foundThe paper's rapeseed phenology dataset (1981–2024, 50 stations) is openly deposited in Science Data Bank under DOI 10.57760/sciencedb.34086, containing structured xlsx tables and diagnostic plots. No custom code was created per the authors.Dataset · publicThe dataset described in this work has been deposited in the Science Data Bank (ScienceDB) under
accession code https://doi.org/10.57760/sciencedb.34086 [27].Open asset ↗Science Data Bank · 10.57760/sciencedb.34086pdf-page:12 lines:1-68Plant phenotyping relevance matchOpenAlex · checked 14 Sept 2026
Published20 Jul 2026International Journal of Agriculture Forestry and Life SciencesCited by 0 · OpenAlex ↗
Selecting ideal drought-tolerant wheat varieties requires a holistic synthesis of digital phenotypes, molecular markers, and agronomic indices. This study evaluated 16 wheat genotypes for drought tolerance by integrating digital phenotyping (UAV-based thermal imaging), molecular data (DREB gene expression profiles), and 12 agronomic indices. While vegetative DREB accumulation remained mostly homogeneous, the generative stage triggered pronounced transcriptional shifts, and late-stage thermal screening revealed highly significant genotypic differences during grain filling. A multivariate PCA biplot identified early canopy temperature differences during tillering (ΔCT_TL) as the most informative non-destructive selection indicator. ΔCT_TL showed a strong positive association with terminal yield stability metrics (YSI and RSI) and a marked negative relationship with the drought sensitivity index (SDI). This early canopy temperature regulation contributed to the maintenance of yield stability in the modern hexaploid variety MFTBY-T and advanced tetraploid lines OR2-T and OR4-S. In contrast, poorly adapted ancient varieties (P5-S and S3-S) exhibited high drought sensitivity accompanied by pronounced late-stage induction of DREB1 and DREB2, suggesting a delayed stress-response mechanism activated under severe tissue dehydration. Conversely, the modern tetraploid variety KZLTN-T and hexaploid landraces appeared to rely on an early vegetative molecular priming strategy. These findings suggest that breeding programs should prioritize the incorporation of vegetative transcriptional traits associated with effective canopy temperature homeostasis into elite genetic backgrounds.
Unmanned aerial vehicle (UAV) photogrammetry offers a cost-effective approach to tree-level detection, however, Structure-from-Motion (SfM) outputs are sensitive to processing choices and site conditions, which can alter canopy representation and reduce individual-tree detection accuracy. Here, we systematically evaluate how SfM reconstruction quality and depth-map filtering influence RGB-only individual-tree detection under controlled acquisition conditions. Objectives were to (i) identify an optimal SfM-derived point-cloud configuration for delineating individual trees, and (ii) implement and test a segmentation workflow (local-maxima treetop detection plus Dalponte2016 in lidR) for detecting and counting trees. We assessed RGB-only SfM for individual-tree detection (ITD) across thirteen 1.21-ha loblolly pine ( Pinus taeda ) plots located in two counties in the state of Alabama in the southeastern United States; eight even-aged plantations and five mixed pine-hardwood stands, while holding image acquisition parameters constant. Using Agisoft Metashape Professional (Agisoft LLC, St. Petersburg, Russia), dense-cloud quality (Lowest, Low, Medium, High, Ultra High) and depth-map filtering (Disabled, Mild, Moderate, Aggressive) were varied in a 5 × 4 full-factorial design; assessment metrics included point-cloud density, canopy-surface completeness, canopy-height-model (CHM) agreement with field heights, and ITD precision/recall/F1. We identified a single high-resolution configuration (Ultra High + Disabled) by screening parameter sets for structural accuracy and suppression of false peaks. Using this configuration, CHMs matched field heights in Washington County, Alabama (R 2 = 0.96; RMSE = 0.44 m; bias = − 0.01 m) and in Cullman County, Alabama (R 2 = 0.44; RMSE = 1.14 m; bias = − 0.09 m); pooled performance was R 2 = 0.98; RMSE = 0.54 m; bias = − 0.01 m. ITD accuracy at the primary 3 m match radius yielded a precision of 0.03; recall = 0.29; F1 = 0.05 in the even-aged plantations (Washington) and a precision of 0.03; recall = 0.12; F1 = 0.05 in mixed pine–hardwood stands (Cullman); pooled F1 = 0.05. The selected parameters and workflow are reproducible and transferable, provide insight into RGB-SfM ITD performance, and indicate when lidar remains preferable for crown delineation.
Reproduction assets foundThe paper's Code availability statement deposits the authors' SfM/ITD processing scripts publicly on OSF (DOI 10.17605/OSF.IO/UXBCZ). Phenotype/field datasets are only available on request, so they are not public assets.Code · publicThe workflow and processing scripts used in this study are publicly available through the Open Science Framework
(OSF) repository: Singh and Narine, [32]. Code Repository for Optimizing SfM Parameters for RGB-Only Individual-Tree
Detection in Loblolly Pine and Mixed Pine-Hardwood Stands. https://doi.org/10.17605/OSF.IO/UXBCZ.Open asset ↗10.17605/OSF.IO/UXBCZpdf-page:12 lines:1-70Plant phenotyping relevance matchCrossref · checked 5 Sept 2026
Accurate field crop counting supports phenotyping, crop monitoring, and yield-related analysis, but real field images often contain scale variation, target overlap, cluttered vegetation, shadows, and visually similar backgrounds. These factors make density-map regression difficult because weak target responses and accumulated false responses in non-target regions can both bias the final count. This study proposes Light-FCCNet, a compact multi-scale framework for accurate field crop counting under a limited parameter budget. The framework integrates three coordinated components: a lightweight feature pyramid aggregation module for compact scale-aware representation, a multi-attention fusion module for refining fused features and suppressing background-induced responses, and an FCC loss that combines pixel-level density regression, count consistency, and structural similarity. Light-FCCNet is evaluated on three public field crop counting datasets, namely GWHD, MTC, and URC. In the canonical ablation trajectory consisting of baseline, baseline_p1, baseline_p1_p2, and full configurations, the full model achieves the lowest errors, with MAE values of 13.28 on GWHD, 18.10 on MTC, and 61.23 on URC, corresponding to reductions of 18.0%, 25.3%, and 35.2% relative to the baseline. Compared with representative general counting baselines and our implementation of TasselNetV2++ under the same experimental protocol, Light-FCCNet obtains the lowest MAE on all three datasets while using only 0.92 M parameters. These results indicate that its advantage comes from the coordinated design of compact pyramid aggregation, attention-guided feature refinement, and counting-oriented supervision, rather than from any isolated module alone.
Abstract Biotic stress is a major, yet under-quantified, driver of global soybean yield losses, and field-based phenotyping under pest pressure remains a critical bottleneck for crop improvement. Using multi-temporal data from soybean genotypes grown under insecticide-protected and unprotected conditions in Brazil, we present a UAV-based, large-scale and non-invasive framework for evaluating genotype performance under natural pest pressure. We introduce a three-dimensional metric that jointly captures productivity, feature-level similarity as a proxy for tolerance, and phenological response through days to maturity. This unified formulation enables field-based quantification of pest resilience and replaces labor-intensive and often unreliable direct pest collection and counting. To operationalize this framework, we integrate vegetation indices and self-supervised visual embeddings into a common representation space linking feature stability, performance response and phenological development. This approach enables robust identification of genotypes that maintain feature integrity, minimize developmental delay and sustain yield under pest pressure, with genotypic differences peaking during the pod-fill (R3–R4) and grain-fill (R5.1–R5.5) stages. Overall, this work establishes a scalable, field-ready paradigm for quantifying crop resilience to biotic stress and provides a practical pathway to accelerate breeding for stable yields under real-world agricultural conditions.
Reproduction assets foundThe paper explicitly states that the analysis code is publicly available in the authors' GitHub repository (jianglong26/soybean-insect-resistance), which directly reproduces this paper's phenotyping pipeline (orthomosaic processing, VI/DINOv3 feature extraction, similarity analysis, genotype ranking). The paper also声明sCode · public540 The code used for analysis is available at https://github.com/jianglong26/Open asset ↗pdf-page:16 lines:1-45Code / dataset availability confirmedEurope PMC · bioRxiv · checked 11 Sept 2026
Solitary bee species that use artificial trap nests are important for agricultural crop production and as indicators of habitat quality. Quantifying cavity-nesting solitary bee foraging and nesting behavior is essential for real-time analysis of population numbers and pollination activity, as well as understanding how environmental conditions shape reproductive success and population dynamics. However, manual observation is labor-intensive, prone to observer bias, and unable to deliver continuous data. Existing automated systems either require individual bee marking or detect presence without resolving nest-tube-level entry and exit events. We developed BeeMonitor, an integrated hardware and computer-vision pipeline that detects nest entry and exit events in cavity-nesting solitary bees from continuous video, using Osmia cornifrons (the horn-faced mason bee) as a model system. A low-cost Raspberry Pi handles solar-powered field recording, while the software combines object detection (YOLOv26), a custom multiple-object tracker (BeeTrack), and a Random Forest classifier trained on trajectory-derived features to distinguish genuine events from incidental detections. Over a 29-day deployment, hardware reliability averaged 97.5% recording coverage. The pipeline achieved 91.3% precision and 87.3% recall (F1 = 0.893), generalizing robustly under leave-one-video-out cross-validation (mean F1 = 0.904). Detected foraging trips correlated strongly with brood cell counts (R2 = 0.849, p < 0.001, n = 19), and a Random Forest model (AUC = 0.820) identified solar radiation as the dominant driver of foraging activity, followed by temperature. BeeMonitor demonstrates that automated computer vision can reliably extract ecologically relevant behavioral data from continuous video, enabling real-time analysis of pollinator behavior and abundance at a temporal and spatial resolution unattainable through manual observation. Its modular design supports adaptation to other species and monitoring contexts.
Reproduction assets foundThe paper explicitly states that source code, 3D STL files, and validation datasets/code are publicly available on the authors' GitHub repository and ScholarSphere. These directly support reproducing the paper's behavioral-event detection pipeline and its evaluation (annotated videos, classifier training/LOVO cross-vaCode · publicSource code for
software and 3D stl files can be found on the official GitHub repository here
https://github.com/Team-Insect-Net/BeeMonitor.Open asset ↗Team-Insect-Net/BeeMonitorpdf-page:2 lines:1-57Dataset · publicValidation datasets and code are available on
Scholars Sphere here
https://scholarsphere.psu.edu/resources/55f1f34b-959f-4c60-8dd3-9b33fb09357f.Open asset ↗Scholars Sphere · 55f1f34b-959f-4c60-8dd3-9b33fb09357fpdf-page:2 lines:1-57Plant phenotyping relevance matchOpenAlex · Europe PMC · bioRxiv · Crossref · checked 5 Sept 2026
Abstract Understanding the link between genetic variation and observable traits is key to crop breeding. Hyperspectral imaging captures physiological and biochemical profiles, but current supervised methods require costly trait annotations and treat each observation as a static snapshot, ignoring the temporal dynamics of plant development. We introduce SST-MAE, a self-supervised framework that learns genotype-discriminative representations from plant hyperspectral developmental trajectories, without requiring phenotypic labels. The model learns to reconstruct masked information, capturing multiple growth trajectories. Validated on 194 field-grown lettuce genotypes across eight time points, the frozen encoder serves as a feature extractor for downstream genotype classification. SST-MAE outperforms raw spectral and linear baselines, achieving AUROC > 0.89 for anthocyanin pigmentation SNPs and 0.77 for leaf serration. The learned features are highly label-efficient, attaining near-full performance with only 30–50% of labeled data, offering a scalable pathway toward high-throughput genetic screening from image-based phenotypes.
Francisco González · Hamza Armghan Noushahi · Danae Hernández · Miguel Garriga · Alejandro del Pozo · Macarena Gerding · Claudia Muñoz‐Espinoza · Salvador A. Gezan · Luis Inostroza
Fall dormancy (FD) and forage yield (FY) are two key traits in alfalfa ( Medicago sativa L.) breeding programs. However, genetic progress has remained limited over the past decades, largely due to the complexity of alfalfa breeding and the reliance on labor-intensive phenotyping methods. High-throughput phenotyping (HTP) using unmanned aerial vehicles (UAVs) represents a promising alternative for rapid and non-destructive crop evaluation. The objectives of this study were to i) estimate FD using UAV-derived canopy height and RGB vegetation indices and ii) evaluate the predictive performance of machine learning (ML) models for FY estimation. A total of 210 alfalfa populations with diverse genetic backgrounds were evaluated over two growing seasons (2023 to 2025) across seven harvests under Mediterranean conditions in central Chile. FY was measured manually, while FD was estimated using both manual and UAV-based approaches. A total of 19 RGB-derived indices (VIs) including plant height (PH) were extracted and used as predictor variables. Five complex predictive ML models were evaluated: PLS, PCR, SVM, ANN, and MLR. The results showed that UAV-derived FD was significantly correlated with FD obtained through conventional methods ( R 2 = 0.88). The automated UAV-based FD phenotyping framework demonstrated slightly higher precision ( R 2 = 0.92) and broad-sense heritability ( H 2 = 0.69) compared to manual measurements ( R 2 = 0.87–0.89; H 2 = 0.64), providing a more reliable selection tool for breeders. Among the tested ML models, SVM and ANN achieved the highest accuracy ( R 2 ≈ 0.73) for FY prediction. These findings demonstrate that integrating low-cost RGB imagery with complex modeling offers a promising avenue that could assist in refining future selection strategies for this genetically complex species.
Sustainable forest management requires inventory workflows that provide spatially explicit structural information while retaining field-based calibration and uncertainty control. This case study evaluated a field-calibrated UAV LiDAR workflow for individual-tree inventory in middle-aged and near-mature larch plantations in Chuanying District, Jilin City, China. UAV laser scanning point-clouds were integrated with six 30 m × 30 m field plots to assess three individual-tree extraction algorithms: point-cloud segmentation (PCS), marker-controlled region growing (MCRG), and region-based hierarchical cross-section analysis (RHCSA). Algorithm performance was evaluated using plot-level recall, precision, F-score, localization RMSE, tree-height and crown-width accuracy, bootstrap confidence intervals, exploratory Wilcoxon signed-rank comparisons, and leave-one-plot-out stability checks. MCRG provided the most balanced numerical performance under the tested configuration, with a mean F-score of 0.845, compared with 0.808 for PCS and 0.827 for RHCSA. However, the MCRG-RHCSA paired difference was not robust across the six plots, and the analysis should be interpreted as a dataset-specific workflow comparison rather than a universal algorithm ranking. Tree height was estimated with comparatively high accuracy, whereas crown-width estimation remained weak, indicating that vertical canopy structure was more reliable than lateral crown delineation. After calibration assessment, the workflow was applied to 157.47 ha of UAV LiDAR survey areas and generated 219,996 algorithm-based detections. These outputs are best interpreted as a spatial decision-support layer for compartment updating, density screening, and field-inspection prioritization, not as an independently verified wall-to-wall stem census.
Yellow rust (YR) is a major threat to both bread and durum wheat production, often causing substantial yield losses. Conventional visual scoring of YR severity, while widely adopted, is labor-intensive, time-consuming, and prone to human error. In this study, we evaluated the predictability (PA), defined as the correlation between predicted and observed values, using genomic and phenomic data for YR severity under multiple prediction scenarios in two biparental wheat populations (bread and durum). YR scoring was conducted on two dates, with YR severity visually assessed while unmanned aerial vehicle (UAV)-based high-throughput phenotyping (HTP) data were collected using a multispectral camera. HTP data were processed to extract spectral wavelengths and vegetation indices (VIs), and all lines were also genotyped using SNP arrays. We tested a diverse set of models, including parametric, machine learning, and deep learning approaches. PA increased markedly when HTP-derived data were used compared with genomic markers alone. For example, support vector regression (SVR) improved from 0.35 (markers only) to 0.87 (wavelengths only). However, integrating genomic and phenomic data did not yield further improvements, as models often plateaued when using HTP-derived features alone. Cross-crop prediction demonstrated promising generalization across bread and durum wheat, achieving PA values up to 0.83. For this last task, best linear unbiased prediction (BLUP) and multilayer perception (MLP) consistently provided robust performance across scenarios. These findings highlight the strong potential of UAV-based HTP for rapid, scalable, and accurate prediction of YR severity in wheat. While genomics retains broad utility for breeding, the practical integration of phenomics and AI-driven prediction pipelines will ultimately depend on breeding program strategies, resources, and objectives.
Reproduction assets foundThe article's Data Availability statement deposits the datasets generated and analyzed in this study (yellow rust phenotyping with UAV spectral data and genomic markers in bread and durum wheat) in the CIMMYT repository under DOI 10.71682/10549375, which is an allowed URL. No author analysis code or trained model is avDataset · publicand scalable strategy for YR assessment in wheat breeding.
Funding
The authors gratefully acknowledge financial support from the Government of Mexico through
the “MasAgro – Cultivos para México” initiative.
Data Availability
The datasets generated and/or analyzed during the current study are available in the CIMMYT
repository: https://doi.org/10.71682/10549375.Acknowledgements
We are deeply grateful to Julio Huerta-Espino for his guidance and support throughout all stages
of this manuscript. We also thank Hedilberto Velásquez Miranda for his valuable assistance with
rust visual score phenotyping, and Neftalí Cruz Pérez for his dedicated support in trial sowing and
field management.
Conflict Open asset ↗10.71682/10549375pdf-raw-page:30 lines:1-37Plant phenotyping relevance matchCrossref · checked 15 Sept 2026
Published17 Jul 2026International Journal of Environment and Climate ChangeCited by 0 · OpenAlex ↗
Apple orchard health monitoring is important for supporting timely crop management under the temperate conditions of Kashmir Valley. The present study evaluated the seasonal behaviour of four Sentinel-2-derived vegetation indices, namely the Normalized Difference Vegetation Index, Soil Adjusted Vegetation Index, Modified Soil Adjusted Vegetation Index and Enhanced Vegetation Index, in a 2 ha apple orchard at SKUAST-K, Shalimar, Kashmir, during the 2020 growing season. Sentinel-2 imagery acquired from April to September was processed using SNAP and QGIS, and mean vegetation index values were extracted for the orchard area. Monthly weather data, including average temperature and rainfall, were obtained from the on-campus meteorological observatory. Ground observations at 20 georeferenced points were used to support the interpretation of canopy development, phenological stage and visible plant health condition. All four vegetation indices showed a seasonal increase from April to July-August, followed by a decline during September-October, corresponding to canopy development, fruit maturation, harvest and senescence. Based on the monthly values presented in the study, temperature showed positive associations with the vegetation indices, with the strongest relationship observed for MSAVI, followed by NDVI and SAVI. Rainfall showed weak and non-significant associations with the indices during the study period. The results indicate that Sentinel-2-derived vegetation indices can reflect seasonal canopy dynamics in apple orchards under the studied conditions. MSAVI appeared particularly useful for representing canopy development, while field observations remained necessary for interpreting pest, disease and phenological effects.
Accurate and reliable diagnosis of grape leaf diseases is essential for sustainable viticulture, enabling timely intervention, reducing economic losses, and supporting precision crop management. Although deep learning (DL) models have demonstrated remarkable classification performance, their reliability under real-world field conditions remains insufficiently explored. In particular, confidence estimates often fail to reflect true predictive correctness when models are exposed to distributional shifts, limiting their practical applicability. To address this challenge, this study proposes a confidence- and uncertainty-aware DL framework for grape leaf disease diagnosis that extends evaluation beyond conventional accuracy-based metrics. Two publicly available grape leaf datasets were employed for stratified five-fold cross-validation in binary and multiclass classification tasks, while a third independently collected dataset was reserved exclusively for leakage-free external validation. EfficientNet-B0 and MobileNetV3-Large were evaluated under four inference configurations: raw inference, temperature scaling (TempScaling), Monte Carlo dropout, and an ensemble strategy combining uncertainty estimation with calibration. External validation demonstrated that both architectures maintained discriminative capability under domain shift. Under ensemble inference, EfficientNet-B0 achieved an accuracy of 73.8%, a macro-F1 score of 73.3%, a Matthews correlation coefficient (MCC) of 0.468, and a receiver operating characteristic area under the curve (ROC-AUC) of 0.804, while MobileNetV3-Large achieved 71.8% accuracy, 71.2% macro-F1, an MCC of 0.429, and a ROC-AUC of 0.787. Despite these promising results, raw predictions exhibited substantial overconfidence, with Expected Calibration Error (ECE) values of 0.287 and 0.312 for EfficientNet-B0 and MobileNetV3-Large, respectively. TempScaling markedly improved calibration quality, reducing ECE to 0.038 and 0.042 without affecting classification performance. Ensemble inference further enhanced the balance between predictive discrimination and reliability. The results demonstrate that strong classification performance alone is insufficient for trustworthy deployment in agricultural environments. Confidence calibration and uncertainty quantification provide complementary information for identifying overconfident predictions and improving decision reliability under field variability. The proposed framework offers a reliability-oriented approach for developing trustworthy artificial intelligence systems for grape leaf disease diagnosis in precision agriculture.
Abstract Purpose of Review Ground-based 3D point cloud technologies, including static terrestrial laser scanning (TLS), mobile laser scanning (MLS), and close-range photogrammetry, are increasingly used for estimation of aboveground vegetation biomass as they provide detailed structural representations across vegetation types; however, a comprehensive synthesis of how point-cloud data are translated into biomass estimates remains lacking. This review evaluates current approaches, performance patterns, and methodological gaps in biomass estimation using 3D ground-based point clouds. Recent Findings We systematically reviewed and analyzed 160 research articles (comprising 171 device-specific studies) published until the end of 2025 (first appearing in 2010). Research was dominated by tree-based applications (74%), with limited attention to shrubs, grasslands or crops. TLS was the prevailing acquisition technology (78%), although MLS adoption is growing. Biomass estimation primarily relied on allometric equations, volume-based reconstructions (e.g., quantitative structure models, voxelizations, convex hull), and parametric regression models. Reported model performance was generally high in tree- and shrub-based studies (median R 2 > 0.8), but more variable in non-woody vegetation types. Despite rapid advances in 3D sensing, point-cloud-native deep-learning approaches remain rarely implemented in biomass estimation workflows. Summary Ground-based 3D sensing is maturing technically, yet methodological heterogeneity persists. Many workflows still depend on destructive calibration data, semi-manual preprocessing, and non-standardized modelling strategies, limiting reproducibility and cross-study comparability. Multi-sensor integration is emerging but lacks consistent upscaling frameworks. Future research should expand coverage of underrepresented vegetation types, promote standardized and automated processing pipelines, and systematically evaluate point-cloud-native deep learning architectures, both for extracting structural proxies and for assessing their capacity to estimate biomass directly.
Rising temperatures and changing weather conditions are accelerating the spread of plant diseases and increasing the threat to global food security. Reliable detection of leaf diseases is therefore essential to protect crop yields and ensure food quality. Deep learning has proven to be a powerful tool for classifying leaf diseases across various crops. Due to the natural variability of plants, plant diseases often appear in irregular structures. Surface unevenness, folds, or dirt particles are common in field images and can be mistakenly identified as important features by convolutional neural networks (CNNs). This is a challenge that has not been sufficiently addressed in previous studies. This study proposes a novel deep learning approach that takes into account both the specific visual characteristics of plant diseases and potential disturbances in the microstructure, such as surface irregularities or prominent leaf veins, which may mislead the model. Using stratified five-fold cross-validation on a peer-reviewed dataset, which comprises 2,801 images of radish leaves across five classes (healthy, three disease classes: mosaic virus, black leaf spot, and downy mildew, and one pest-affected class: flea beetle), the proposed method achieved an average and balanced accuracy of 99.86%, establishing a new dataset-level benchmark in the field and demonstrating its effectiveness. The results indicate that the proposed approach may provide a promising basis for future applications in agricultural field monitoring, automated sorting and post-harvest quality control, offering potential to reduce both food waste and associated costs.
Pavel Klimeš · Vladimír Voral · Nabila M. Gómez Mansur · Jakub Vašák · Jana Kholová · Sanja Ćavar Zeljkovıć · Monika Rozehnalová · Lukáš Spíchal · Jan Masner · Nuria De Diego
Early crop establishment strongly influences plant performance and yield, making seedling emergence an important trait in crop phenotyping, breeding, and stress physiology studies. However, emergence monitoring is still commonly performed manually and typically records only the final emergence percentage, limiting the analysis to other dynamic observations. Automated image-based approaches are promising but remain challenging due to the small size of plant structures, heterogeneous soil backgrounds, and variability across imaging systems. Here, we present SPROUT (AI- based S eedling PR edicti O n and trait extraction U sing RGB T ime-series), a low-cost RGB imaging pipeline for automated prediction of crop emergence dynamics and trait extraction. The system integrates instance segmentation, object-detection–based data reduction, and temporal deep learning to estimate the emergence time of individual seedlings from RGB image sequences. The pipeline then automatically reconstructs emergence curves and extracts associated traits, including final emergence percentage, EC50, and emergence synchronicity. SPROUT was developed and evaluated using barley and wheat datasets acquired with different RGB cameras under controlled growth-chamber conditions. In the development and retraining settings, the best-performing TCN model achieved 90.0% per-well accuracy with a ± 2h tolerance, supporting accurate emergence curve reconstruction. In an independent inference-only dataset, the model still captured approximate emergence dynamics, although accuracy decreased to 59.3%, indicating that SPROUT is best used as a modular pipeline that can be retrained or fine-tuned for new crop, camera, or experimental domains. A cadmium-stress case study in two contrasting wheat genotypes showed that SPROUT-derived traits captured genotype-specific establishment strategies associated with growth and metabolic responses.
Reproduction assets foundThe paper's SPROUT emergence-prediction pipeline code is publicly available on GitHub with explicit availability language, and raw images plus morphology/metabolic data are deposited on Zenodo (10.5281/zenodo.18889863). The GitHub URL is in allowed_urls; the Zenodo DOI is not, so only the code asset is listed as an ad-Code · publiccan be found online at https://doi.org/10.1016/j.compag.2026.112184.Data availability
The raw images and raw data for the morphology and metabolic
profiling on the case study are available in ZENODO (10.5281/zen
odo.18889863), and the code for the machine learning pipeline and
emergence curve analysis are available on GitHub (https://github.com/kit-pef-czu-cz/sprout-emergence-prediction).References
Albarenque, S., Basso, B., Davidson, O., Maestrini, B., Melchiori, R., 2023. Plant
emergence and maize (Zea mays L.) yield across multiple farmers’ fields. Field Crops
Res. 302. https://doi.org/10.1016/j.fcr.2023.109090.Arsovski, A.A., Galstyan, A., Guseman, J.M., Nemhauser, J.L., 2012.
PhotomorpOpen asset ↗kit-pef-czu-cz/sprout-emergence-predictionpdf-raw-page:13 lines:78-112Code / dataset availability confirmedOpenAlex · Europe PMC · checked 5 Sept 2026
Orthoimagery derived from unmanned aerial vehicles (UAVs) has become a valuable data source for crop-growth monitoring. Individual plant-level (IPL) information enables high-throughput analyses by capturing plant-to-plant variability within fields. However, reliable IPL-based analysis requires accurate extraction of plant-specific regions, which remains challenging in soybean cultivation due to weed interference and canopy overlap. This study proposed an automatic preprocessing framework for IPL soybean growth monitoring that integrates deep-learning-based semantic segmentation with a furrow-guided region of interest (ROI) generation strategy using UAV imagery. A segmentation model was developed using combinations of RGB and multispectral orthoimagery, and a furrow line detection algorithm was designed to generate IPL ROIs aligned with crop rows. The ensemble model combining U-Net, DeepLabV3+, and SegFormer achieved the most stable performance (F1-score up to 0.94 and IoU up to 0.89). The furrow-guided ROI generation algorithm also accurately estimated crop counts, showing strong agreement with manual observations (R² = 0.90 and RMSE = 6.35). The generated IPL ROIs enabled accurate quantification of growth-related features, with strong agreement between automatically generated and manually delineated ROIs (R² > 0.90). Overall, the proposed preprocessing framework provides a practical and scalable solution for UAV-based high-throughput phenotyping in soybean and other ridge-based cropping systems.
Reproduction assets foundThe authors state that the complete implementation of their IPL soybean preprocessing pipeline (semantic segmentation + furrow line detection) is publicly available on Zenodo. The annotated sample dataset, however, is only available upon request from the corresponding author, so it is not a public asset.Code · publicThe complete implementation of
this pipeline is publicly available at https://doi.org/10.5281/zenodo.21095307.Open asset ↗zenodo · 10.5281/zenodo.21095307pdf-page:7 lines:1-62Plant phenotyping relevance matchCrossref · checked 15 Sept 2026
Reliable ecological indicators of mangrove structure and carbon storage are essential for monitoring coastal ecosystem conditions, yet their accuracy remains uncertain in tall, structurally heterogeneous forests, where Earth observation products differ in sensor physics, spatial resolution, and acquisition dates. Here, we present a multi-scale framework to evaluate, calibrate, and improve two widely used ecological indicators of mangrove condition—canopy height and aboveground biomass (AGB)—across approximately 7000 ha of mangroves in the Marapanim estuary, northern Brazil. The framework integrates UAV photogrammetry, radar-derived digital elevation models (TanDEM-X and SRTM), and field measurements to quantify cross-scale discrepancies and identify the main sources of uncertainty affecting indicator retrieval. High-resolution UAV canopy-height models revealed exceptionally tall Avicennia forests reaching up to 53 m, among the tallest mangroves reported globally. At the local scale, mean AGB reached approximately 648 Mg ha −1 in the southern Avicennia -dominated sector and 430 Mg ha −1 in the northern mixed Rhizophora–Avicennia sector, with local maxima of ∼800 Mg ha −1 . In contrast, radar-derived products yielded substantially lower estimates of canopy height and biomass, with height differences of 8–10 m in tall and structurally heterogeneous stands. These discrepancies reflect the combined effects of sensor-dependent canopy representation, spatial averaging, and temporal mismatch between historical radar acquisitions and recent UAV observations. To improve the ecological interpretation of these products, we implemented a calibration strategy linking field and UAV measurements to satellite observations and complemented it with UAV-based three-dimensional volumetric reconstruction of individual trees as an independent structural check on allometric biomass estimates. Our results show that canopy height and AGB derived from coarse-resolution radar products can systematically underestimate mangrove structural condition and carbon storage in tall forests unless locally calibrated. Beyond documenting exceptionally tall and carbon-dense Amazonian mangroves, this study provides a transferable framework for evaluating and improving ecological indicators of forest structure and biomass in complex coastal ecosystems.
Whole plant leaf expansion (shoot expansion) under drought drives the trade-off between water saving for later grain production and canopy photosynthesis. Fine-tuning shoot expansion could therefore become a target of genetic progress for drought-prone environments. However, its components (axis production, i.e. tillering, leaf production on each axis, and individual leaf elongation) may have their own genotypic variability and plasticity under drought, making hard to calibrate crop simulation models and specify breeding targets. In this study, we focused on the genetic diversity of bread wheat and durum wheat to determine the links and trade-offs between the underlying processes of shoot expansion under drought and how it translates at the whole plant and canopy level. For that, we used non-destructive imaging both in the field and controlled condition platforms to determine their dynamics and analyze their relative contribution to the genotypic variability of whole-plant shoot expansion under drought. Results show that shoot expansion measured at plant level in controlled environment was associated with that measured at canopy level in the field, indicating that controlled phenotyping platforms can capture the genotypic variability of growth in the field. Both whole-plant and canopy expansion were associated with tillering rate. In addition, the sensitivity of shoot growth and tillering to soil water deficit were correlated, indicating that both tillering ability and sensitivity to water deficit drive the genotypic variability of shoot expansion. Overall, dissecting shoot- expansion dynamics allowed determining the links between shoot expansion traits under drought, and provides key targets in phenotyping, modelling and breeding for drought environments.
Aims Climate change is altering northern peatland plant communities, shifting from Sphagnum mosses to vascular plants. This transition impacts ecological functions like carbon sequestration, making long-term vegetation monitoring at the site scale more critical than ever. However, current monitoring methods tend to focus on specific species or functional groups with limited spatial coverage. This study uses remote sensing to infer the spatial structure and temporal variations of peatland plant communities. Location Temperate peatland in Pyrenees Mountains, France (Bernadouze, Vicdessos). Methods Nine plots were selected across diverse microhabitats and sampled three times over the growing season of 2023 (May, June, and July). Plant species abundances were recorded, and 45 vegetation indices were derived from drone and Sentinel-2 multispectral imagery. Five vegetation indices were selected to fit a joint Species Distribution Model (JSDM) and a Random Forests (RF) model, and map species spatial distribution. Principal Coordinates Analysis (PCoA) identified plant community composition, and spatiotemporal variations were quantified in relation to environmental variables. Results Plant species occurrences could be predicted from multispectral imagery using the JSDM, with drone-based inferences (mean R 2 = 0.36) outperforming Sentinel-2 (mean R 2 = 0.29). Model performance was high for abundant species ( R 2 > 0.5), whereas predictions for rare species were less accurate ( R 2 R 2 > 0.65, P R 2 = 0.40; P R 2 = 0.04; P Conclusion This study demonstrates that drone multispectral imagery can be used to predict peatland vegetation richness and community composition and capture fine-scale heterogeneity in a small and fragmented peatland site, outperforming satellite data in spatial precision. Although our model was less accurate using satellite imagery, the use of Sentinel-2 imagery enabled long-term community tracking. By combining both, our predictive modelling framework provides a promising preliminary tool to monitor climate-induced shifts in species distributions, supporting targeted conservation.
Reproduction assets found保存済みの本文根拠を更新済みルールで再検証し、公開資産1件を確認しました。Code · publicCodes to replicate main analyses are available at https://github.com/vjassey/peatland_vegetation_mapping .Open asset ↗vjassey/peatland_vegetation_mappinglines:369-375Plant phenotyping relevance matchEurope PMC · checked 5 Sept 2026
Florida and California produce 98% of U.S. strawberries, with Florida growers' profitability depending on high yields early in the season (November-January), when prices are the highest in the U.S. market. This study aims to model the cumulative Marketable Yield and Delta Yield curves (difference in cumulative Marketable Yield between consecutive harvest time points) of strawberry genotypes to facilitate selection for greater early season productivity. The dataset comprised thirteen seasons (2013-14 to 2025-26) of advanced selection trial data. Marketable Yield trajectories were modeled using Legendre polynomial smoothing, with optimal degree selection balancing flexibility and noise reduction. The resulting coefficients served as surrogate phenotypes for genomic prediction. Forward prediction cross-validation was implemented for five seasons (2021-22, 2022-23, 2023-24, 2024-25, and 2025-26), with each season predicted using the information from all preceding seasons. For cumulative yield, across all five seasons, reconstructed curves from the Legendre models presented a clear temporal trend, with predictive ability increasing from low early-season values to peaks around Trait Dates (weeks) 7-8. In the 2021-22 and 2022-23 seasons, Legendre models showed higher predictive ability than single time-point predictions but were comparable to single-time point predictions for the other seasons. Legendre polynomial models utilizing Delta Yield achieved moderate predictive ability across five validation seasons, with consistent advantages over single time point models particularly in earlier seasons, indicating that genetic control extends beyond total yield to the trajectory of yield accumulation. Overall, Legendre modeling effectively captured the temporal dynamics of yield development while describing the trajectory with only a few parameters.
Semako Ibrahim Bonou · Rosana Araujo Martins Lucena · Agda Malany Forte de Oliveira · Igor Eneas Cavalcante · Tulio William da Silva Gonçalves · Priscylla Marques de Oliveira Viana · Guilherme Felix Dias · Rener Luciano de Souza Ferraz · André Allison Rodrigues da Silva · José Dantas Neto · Carlos Alberto Vieira de Azevedo · Alberto Soares de Melo
Abstract Purpose The variability in tolerance to water stress among cowpea genotypes requires fast and accurate phenotyping methods. The integration of infrared thermography with artificial intelligence is emerging as a robust solution for large-scale, non-invasive monitoring. Thus, the objective was to train models to identify genotypes and diagnose water stress in cowpea using artificial intelligence algorithms to process infrared thermographic images. Methods Ten genotypes (five varieties: Corujinha – G1, Paulistinha – G2, Sempre Verde – G3, Pintado – G4, and Rabo de Tatu – G5) and the cultivars BRS Novaera – G6, BRS Pajeú – G7, IPA 206 – G8, BRS Tapaihum – G9, and BRS Miranda – G10) were subjected to four water regimes (25%, 50%, 75%, and 100% of ETc). Thermographic images were collected at the V3 and R2 stages and processed using Deep Learning architectures (InceptionV3, SqueezeNet, VGG16, and VGG19) to extract features (vectorization). The k-NN, Decision Tree, Random Forest, SVM, Neural Network, and AdaBoost algorithms were trained to classify stress levels and genotypes. Results The vegetative stage (V3) proved more effective for diagnosis than the reproductive stage (R2), exhibiting more stable thermal signatures. The SVM algorithm, combined with the VGG16 vectorizer, achieved the best performance, achieving an accuracy greater than 0.910 in classifying water regimes. The landrace varieties exhibited thermal patterns distinct from those of the improved cultivars, enabling high-precision genotypic identification. Conclusions The proposed approach demonstrates that infrared thermography, combined with machine learning models, is an effective tool for high-throughput digital phenotyping, optimizing the selection of drought-tolerant materials and irrigation management in precision agriculture.
Fusarium head blight (FHB) is a major mycotoxigenic disease of wheat, causing yield and quality losses and deoxynivalenol contamination. Rapid, non-destructive tools are needed to detect FHB, monitor wheat physiological responses, and evaluate sustainable management strategies, including biological control agents. Although vegetation spectroscopy is widely used for high-throughput phenotyping, most spectral studies focus on binary disease detection, while the capacity of hyperspectral data to capture concurrent host–pathogen–biocontrol responses across leaf and canopy scales remains underexplored. Here, we tested a full-range (400–2400 nm) hyperspectral phenotyping framework to track early interactions among winter wheat, FHB, and Trichoderma gamsii T6085. Two cultivars, Bingo and Rebelde, with higher and lower FHB susceptibility, respectively, were treated with a chemical fungicide (Chem) or T. gamsii T6085 (Bioc) under FHB pressure. Leaf- and canopy-level spectra were acquired at 2, 5, and 14 days post-inoculation, alongside gas exchange, water status, and chlorophyll measurements. Permutational multivariate analysis of variance (PERMANOVA) tested whole-spectrum effects, partial least squares discriminant analysis (PLS-DA) explored class separability, and partial least squares regression (PLSR) estimated physiological traits. PERMANOVA detected genotype × inoculation × treatment interactions from 5 days post-inoculation at leaf and canopy levels. PLS-DA revealed treatment- and cultivar-dependent spectral fingerprints, but overall low-to-fair validation performance indicates that these class-specific patterns should be interpreted as exploratory and not as evidence of operational treatment discrimination. PLSR provided high accuracy for chlorophyll content and osmotic potential, moderate accuracy for CO 2 -assimilation traits, and poor accuracy for transpiration and leaf water potential. While the workflow is scalable as an experimental and analytical framework, its operational deployment will require broader validation across sites, seasons, cultivars, disease-pressure conditions, and sensing platforms.
Eggplant (Solanum melongena) is a major cash crop, yet field detection of its pests and diseases remains difficult because disease evidence is simultaneously occluded by foliage, blurred at lesion boundaries, and highly variable in scale. Fruit rot is especially challenging: the waxy epidermis and purple anthocyanin-rich surface reduce chromatic contrast, while infection spreads gradually from water-soaked tissue to necrotic tissue, producing diffuse borders between diseased and healthy regions. In this work, we reframe eggplant disease detection through a context-boundary-scale coupling principle, which states that accurate field detection should jointly model incomplete contextual cues, ambiguous lesion boundaries, and scale-varying symptom morphology rather than optimize these cues independently. ISA-YOLO is proposed as an implementation of this principle on top of YOLOv13 through coordinated context modeling, boundary-aware aggregation, and progressive multi-scale fusion. Experiments on two public datasets show that ISA-YOLO achieves 78.1 and 77.7% mAP at 30.66 and 31.74 FPS, outperforming mainstream detectors in overall trade-off between accuracy and speed. After pruning and quantization, inference speed increases to about 75 FPS while maintaining strong accuracy. These results indicate that the proposed principle provides an effective pathway for accurate and deployable eggplant pest and disease detection in smart agriculture.
Reproduction assets foundThe paper uses four public Roboflow image datasets (BISU, UTM, Papaya, Tomato) and states that supporting data and code are publicly available on Zenodo, all with explicit URLs in the Data availability section.Dataset · publicwas supported by National Natural Science Foundation of China grants (62472269, 62072291).
Data availability
This research relies entirely on four publicly available detection datasets: Papaya ( https://universe.roboflow.com/hm-xsfz9/papaya-eswlk-r2ydy ), Tomato ( https://universe.roboflow.com/test-fqgof/tomato-rotten ), BISU ( https://universe.roboflow.com/bohol-island-state-university-vgjlb/eggplant-disease-detection ), and UTM ( https://universe.roboflow.com/utm-xpfqs/eggplant-disease-detection-5fuqv ). The authors believe that the creators of these public datasets complied with relevant institutional, national, and international guidelines and legislation when collecting the data. No newOpen asset ↗eggplant-disease-detectionlines:1509-1560Dataset · publicty
This research relies entirely on four publicly available detection datasets: Papaya ( https://universe.roboflow.com/hm-xsfz9/papaya-eswlk-r2ydy ), Tomato ( https://universe.roboflow.com/test-fqgof/tomato-rotten ), BISU ( https://universe.roboflow.com/bohol-island-state-university-vgjlb/eggplant-disease-detection ), and UTM ( https://universe.roboflow.com/utm-xpfqs/eggplant-disease-detection-5fuqv ). The authors believe that the creators of these public datasets complied with relevant institutional, national, and international guidelines and legislation when collecting the data. No new data collection was performed for this research. The authors confirm that the use of these datasets in thOpen asset ↗eggplant-disease-detection-5fuqvlines:1509-1560Dataset · publicof the outcomes of the Provincial Undergraduate Training Program on Innovation and Entrepreneurship (Number: S202510108100). This work was supported by National Natural Science Foundation of China grants (62472269, 62072291).
Data availability
This research relies entirely on four publicly available detection datasets: Papaya ( https://universe.roboflow.com/hm-xsfz9/papaya-eswlk-r2ydy ), Tomato ( https://universe.roboflow.com/test-fqgof/tomato-rotten ), BISU ( https://universe.roboflow.com/bohol-island-state-university-vgjlb/eggplant-disease-detection ), and UTM ( https://universe.roboflow.com/utm-xpfqs/eggplant-disease-detection-5fuqv ). The authors believe that the creators of these publicOpen asset ↗papaya-eswlk-r2ydylines:1509-1560Dataset · publicnovation and Entrepreneurship (Number: S202510108100). This work was supported by National Natural Science Foundation of China grants (62472269, 62072291).
Data availability
This research relies entirely on four publicly available detection datasets: Papaya ( https://universe.roboflow.com/hm-xsfz9/papaya-eswlk-r2ydy ), Tomato ( https://universe.roboflow.com/test-fqgof/tomato-rotten ), BISU ( https://universe.roboflow.com/bohol-island-state-university-vgjlb/eggplant-disease-detection ), and UTM ( https://universe.roboflow.com/utm-xpfqs/eggplant-disease-detection-5fuqv ). The authors believe that the creators of these public datasets complied with relevant institutional, national, and internatOpen asset ↗tomato-rottenlines:1509-1560Code · publicarch. The authors confirm that the use of these datasets in this study is fully compliant with their original licenses and ethical guidelines. The final images presented in the article accurately reflect the original data and meet community standards. The data and code supporting the conclusions of this article are available at https://zenodo.org/records/19425300 .
Declarations
Ethics approval and consent to participateOpen asset ↗Zenodo · 19425300lines:1509-1560Plant phenotyping relevance matchEurope PMC · checked 5 Sept 2026
Abstract Accurate plot-level sugarcane yield forecasting is essential for optimizing agricultural management, resource allocation, and operational planning. Existing forecasting approaches are often limited by their inability to capture temporal crop dynamics and local biophysical variability, reducing their usefulness for real-time decision-making. To develop and evaluate a Machine Learning (ML)-based framework for short-term sugarcane yield forecasting at plot level using age-segmented Normalized Difference Vegetation Index (NDVI) derived from Sentinel-2 imagery, and to determine the earliest crop stage at which reliable yield predictions can be obtained. An integrated dataset was constructed by combining productivity records from 2,132 sugarcane plots across six harvest seasons (2016–17 to 2021–22) with NDVI time series derived from Sentinel-2 satellite imagery. NDVI observations were aggregated into phenology-based temporal intervals, from which statistical features were extracted. Ten ML regression algorithms were evaluated under two forecasting schemes: a global model trained with all observations and a sector-specific approach that developed localized models for individual production sectors. Model performance was assessed using RMSE and R² on an independent test set. The sector-specific approach outperformed the global model, achieving an RMSE of 12.48 TCH and an R² of 0.7840 on the independent test set, compared with an RMSE of 16.75 TCH and an R² of 0.5724 for the global model. Sparse Partial Least Squares (spls) and Support Vector Machines with Polynomial Kernel (svmPoly) were the most frequently selected algorithms. SHAP analysis revealed that Median NDVI was the dominant predictive feature, while the Elongation I stage was the most influential phenological period. Reliable forecasts were obtained from the fifth month of crop growth (RMSE = 14.13 TCH), and prediction accuracy improved progressively as the crop matured. The proposed framework also surpassed traditional expert estimations (RMSE = 15.47), providing earlier and more accurate yield forecasts. This study demonstrates that localized, sector-specific ML models combined with temporal NDVI dynamics can provide accurate and operationally useful plot-level sugarcane yield forecasts. The framework supports proactive agronomic management, improves planning and budgeting processes, and offers a scalable methodology for precision agriculture and sustainable sugarcane production systems.
Plant diseases lead to substantial yield losses and pose a persistent threat to global food security, creating an urgent demand for high-throughput, accurate, scalable, and non-destructive disease monitoring approaches. Remote sensing has emerged as a powerful tool, yet progress in plant disease detection remains fragmented across various disciplines, tasks, sensing methods, and data modalities. This review introduces a hex- view perspective to synthesise remote sensing–based plant disease detection within a cohesive conceptual framework. Instead of treating sensing technologies, algorithms, and datasets independently, the hex-view incorporates six interconnected dimensions that jointly capture how biological processes, measurement scale, and data characteristics constrain disease detectability, including when detection is possible and how reliably it can be achieved. The hex-view framework comprises six interconnected dimensions and forms an integrated framework called BTSCAD: (1) Biology (B): Plant-pathogen interactions constituting the biological foundation of disease development and expression. (2) Task (T): The diverse disease detection tasks and their corresponding research objectives. (3) Sensor (S): The sensing modalities that define the data acquisition type and richness of captured information. (4) Condition (C): The environmental conditions, sensing platforms, and spatial scales that shape disease observations and bridge controlled experiments and real-world deployment across leaf, canopy, plot and regional scales. (5) Algorithm (A): The classical and state-of-the-art data analysis algorithms used to extract disease-related information from sensor data. (6) Dataset (D): The data sources that underpin model development, evaluation, and generalisability. The hex-view perspective provides a clear framework for interpreting previous research and identifying future research directions. This review lays a structured foundation for developing robust, interpretable, and transferable disease detection systems, supporting advancements in precision agriculture, high-throughput phenotyping, and sustainable crop production.
Introduction Early-stage detection and classification of lettuce heat responses are essential for non-destructive phenotyping, yet conventional assessment mainly relies on visible symptoms and manual observation. Methods This study constructed a lettuce hyperspectral dataset comprising heat-sensitive and heat-tolerant varieties under control and high-temperature treatments, and proposed the Dynamic Selective Peak Transformer (DSPformer). DSPformer integrates edge-enhanced feature extraction, dynamic multi-scale spatial-spectral representation, Peak-k selective attention, and a confusion-aware dynamic focal loss to enhance discriminative features while reducing spectral redundancy, class imbalance, and inter-class confusion. Results Under the patch-level evaluation protocol, DSPformer achieved 96.22% accuracy, 95.55% recall, 96.35% precision, and 95.95% F1-score, outperforming the compared CNN- and Transformer-based models. Day-wise evaluation showed that DSPformer reached 82.61% accuracy on Day 1 and 96.55% on Day 3, before visible heat-stress symptoms appeared on Day 6. Under a plant-level partition protocol, DSPformer maintained robust performance with 93.76 +/- 0.49% accuracy. Additional evaluation on the Indian Pines benchmark further demonstrated the applicability of DSPformer to general hyperspectral image classification. Discussion These findings suggest that hyperspectral imaging can capture heat-stress-sensitive information beyond visual phenotypes, and that DSPformer provides a promising framework for early, non-destructive lettuce heat-response screening and hyperspectral phenotyping-assisted breeding.
Accurate yield prediction for major grain and oilseed crops, including soybean, corn, wheat, and rice, is essential for food-security assessment and precision field management. This study presents a structured integrative review of UAV-based crop yield prediction and follows PRISMA-guided procedures for literature search, screening, and evidence synthesis. Seventy peer-reviewed studies published between 2018 and 2025 were synthesized within a "Data-Ground Truth-Model-Decision" framework. Beyond summarizing UAV platforms, sensor configurations, feature-engineering strategies, and model architectures, the review explicitly distinguishes among microplot, field, and regional prediction scales, and evaluates the characteristics and limitations of yield-label acquisition methods, including manual harvest, plot-combine harvest, and combine yield-monitor data. Existing evidence indicates that the reliability of UAV-based yield prediction depends not only on optimal image acquisition windows, multi-source feature fusion, and model architecture, but also on scale-consistent yield labels, spatially aware validation strategies, and clearly defined model outputs, such as plot-level scalar yield, field-scale yield maps, and regional yield estimates. Major bottlenecks include scale mismatch between UAV imagery and yield labels, error propagation during yield-map generation, limited cross-year and cross-region transferability, weak causal interpretability, and difficulties in deploying models under complex operational field conditions. Future research should emphasize scale-explicit benchmark datasets, quality-controlled ground-truth yield acquisition, UAV-satellite-ground data fusion, spatiotemporal deep learning, and edge-cloud collaborative systems that can translate prediction outputs into agronomic decisions. This review provides a practical pathway for developing robust, interpretable, and deployable UAV-based yield prediction systems for major grain and oilseed crops.
India’s economy is primarily based on agriculture. Agriculture has significant contribution in nation’s GDP. Food security and employment significantly influenced by agriculture. However factors like uncertain weather conditions, poor quality of seeds and plant diseases impact on agriculture productivity. Computer vision and DL algorithms are most crucial components of precision agriculture. Early detection can improve decision making, maximize pesticide use, and preserve harvests. Using CNN architectures, segmentation-based approaches, handcrafted feature-based methods, and hybrid approaches incorporating Machine Learning and Deep Learning this study seek to provide review of recent publications from 2020 to 2026. The review was carried out using a variety of publications with different datasets, methodologies, and outcomes. The findings show that DL, especially CNN and transfer learning models, performed better than machine learning techniques. It points out several significant problems, such as dataset imbalance, insufficient generalization, computing inefficiency, and a dearth of real-world data. Future research topics are also suggested which includes IoT-driven real-time solutions, lightweight architecture, domain adaption, and multimodal imaging. This review aims to develop plant disease detection technologies that are more dependable, scalable, and field deployable.
Artificial intelligence (AI) has emerged as a transformative tool for plant health monitoring, offering new opportunities for scalable, timely, and data-driven pest and disease management in agriculture. This review provides a comprehensive synthesis of AI-based methods for pest and plant disease detection, systematically organizing existing literature across sensing modalities, learning paradigms, and deployment scales. We distinguish between population-level pest monitoring, plant-centric visual inspection, and field-scale surveillance, as well as between post-symptomatic disease recognition and pre-symptomatic detection enabled by spectral imaging technologies. Beyond summarizing recent advances, this work places strong emphasis on critical analysis, discussing fundamental limitations related to data scarcity, domain shift, generalization under field conditions, and the challenge of disentangling biotic from abiotic stress factors. The review further examines the distinction between correlation-driven AI predictions and causal disease understanding, positioning AI as a complementary decision-support tool alongside established diagnostic methods. Building on these insights, we outline key future research directions, including multimodal sensor fusion, explainable and trustworthy AI, edge-based deployment for real-time monitoring, and the development of foundation models for unified agricultural intelligence. This review aims to serve as both an accessible entry point and a critical reference for advancing AI-driven plant health management.
Abstract Genomic selection has accelerated genetic gain in many breeding programs worldwide but genotype-by-environment-by-management (GxExM) hampers further progress for systems where these interactions are important and not well represented in the training data. Process-based crop growth models (CGMs), which encode physiological relationships between plants and their environments, can extrapolate to novel conditions but cannot directly leverage genomic information. Coupling these complementary approaches in the Crop Growth Model–Whole Genome Prediction (CGM-WGP) framework addresses both limitations, yet applications in horticultural crops remain scarce. In this study, we apply CGM-WGP to predict flowering time in broccoli ( Brassica oleracea var. italica ) and common bean ( Phaseolus vulgaris L.), two horticultural species with contrasting physiological responses to temperature and photoperiod. Genotype-specific thermal time requirements and photoperiod parameters were jointly estimated with genome-wide marker effects and predictions were compared against a Reaction Norm Genomic Best Linear Unbiased Prediction (RN-GBLUP) benchmark across four cross-validation scenarios of increasing predictive difficulty. RN-GBLUP achieved the highest accuracy under sparse-testing scenarios where training data covered all target environments, while CGM-WGP outperformed RN-GBLUP when predicting untested environments and untested genotype-environment combinations (broccoli: Pearson r = 0.66, RMSE = 9.4 days; bean: r = 0.86, RMSE = 5.2 days). These results demonstrate that CGM-WGP can be applied to horticultural crops using genome-wide markers alone, but without requiring prior identification of quantitative trait loci. CGM-WGP also provides a modular foundation that can be extended to predict the timing of other developmental transitions and output traits such as biomass and yield.
The genetic identity of coffee cultivars is fundamental to the specialty coffee sector, where premium prices are paid under the assumption that the purchased planting material corresponds to the declared variety. However, many producing countries lack the certification infrastructure necessary to guarantee this identity in their informal seed systems, exposing producers to undetected varietal non-conformity. In this study, we examine a case from a specialty coffee ( Coffea arabica L.) farm in southern Ecuador where seeds labeled as Sidra (USD 100/kg) and Gesha (USD 500/kg) were purchased without genetic or phytosanitary certification. Using a combination of SSR-based DNA fingerprinting and quantitative morphological characterization, including plant architecture, leaf functional traits, and fruit characteristics, we documented varietal identity and assessed the discriminant capacity of morphological traits across the four resulting morphotypes. Using eleven microsatellite markers for SSR fingerprinting, we found that two of the four morphotypes did not match their declared commercial identity. One plant sold as Sidra was identified as compatible with Batian, a composite variety of Kenyan origin that is genetically unrelated to Ethiopian landraces. The plants acquired as Gesha corresponded to a pure Ethiopian landrace that is genetically similar to, but not identical to, the Panamanian Geisha reference accession T.02722. Only two morphotypes were confirmed as Sidra. Furthermore, the placement of Sidra within the Core Ethiopia genetic group is consistent with prior population-level analyses and with its likely status as a selected Ethiopian landrace rather than a variety of hybrid origin. Morphological linear discriminant analysis achieved 82.4% overall classification accuracy under leave-one-out cross-validation (LOOCV), with internode length dominating the first discriminant function (LD1 = 66.6%). These results demonstrate that varietal nonconformity in the specialty coffee seed sector can extend to the inadvertent introduction of genetically unrelated material and underscore the urgent need for accessible seed certification.
Reproduction assets foundThe paper's morphological/functional trait dataset (used for the phenotyping and LDA analysis) is explicitly stated to be publicly available on Figshare (10.6084/m9.figshare.32841344). No author analysis code repository is stated; other URLs in the text are generic libraries or cited prior work.Dataset · publicThe morphological and functional trait dataset generated and analyzed in this study is publicly available in the Figshare repository at 10.6084/m9.figshare.32841344 .Open asset ↗Figshare · 10.6084/m9.figshare.32841344lines:526-568Plant phenotyping relevance matchCrossref · Europe PMC · checked 5 Sept 2026
Abstract This paper presents a novel medium-adaptive wideband near-field antenna for early detection of internal cavities in plant stems, including branches and small trunks. Unlike conventional antennas designed for free-space operation, the proposed antenna is explicitly engineered to operate in close proximity to a lossy, anisotropic, and dispersive cylindrical medium representing wood tissues. A physics-based electromagnetic model of the stem is incorporated into the design process, enabling accurate optimization under realistic dielectric loading conditions. The antenna consists of a compact quasi-planar dipole with blended arms integrated with a medium-adaptive dual-ring balun that ensures balanced current excitation and stable impedance matching under strong near-field loading. Both simulation and experimental measurements demonstrate wideband impedance matching to a 50 Ω source over the 2.0–3.0 GHz frequency range. Surface current distribution and specific absorption rate (SAR) analyses confirm efficient electromagnetic coupling into the stem tissues with minimal radiation leakage. To evaluate the sensing capability of the proposed design, a conceptual two-element antenna system is introduced as a feasibility study for cavity detection. The detection performance is assessed through a sensitivity-driven framework based on variations in both self- and mutual-scattering parameters. A comprehensive sensitivity analysis is conducted to quantify the response of the system to changes in cavity diameter, radial position, and angular location. The results demonstrate that while the reflection coefficient is primarily sensitive to near-surface inhomogeneities, the mutual coupling between antenna elements provides strong and reliable sensitivity to internal cavity characteristics. Based on the sensitivity analysis, an optimal operating frequency band centered at 2.76 GHz and an appropriate antenna clearance are identified to maximize detection performance. The proposed antenna and sensing methodology are further validated through experimental measurements, confirming the consistency with numerical results. Simulation results demonstrate cavity-detection sensitivity, while experimental measurements validate the antenna impedance matching and mutual-coupling characteristics. The compact geometry of the antenna enables scalable multi-element configurations, establishing a practical framework for non-destructive, microwave-based monitoring of internal tree degradation in agricultural and forestry applications.
ABSTRACT The common bean is vital for food security, but its productivity is often limited by competition with weeds, requiring the use of herbicides. The response of genotypes to herbicides such as fomesafen and imazamox is variable, and the traditional evaluation of phytotoxicity through visual methods is subjective. Therefore, the present study aimed to: (i) propose a methodology based on image analysis for phenotyping herbicide‐induced phytotoxicity in common bean genotypes, aiming to reduce the subjectivity of traditional visual assessments; (ii) characterize common bean genotypes under the effects of different herbicides and their doses in progenies and parental lines, based on morphophysiological traits and indices derived from visible RGB (red, green, and blue) digital images. The experiment was conducted in a completely randomized design under a 3 × 3 × 6 factorial scheme (herbicide × dose × genotype) with three replications, evaluating fomesafen and imazamox at doses of 0%, 100%, and 200% of the recommended rates. Data were collected on visual phytotoxicity, plant height, stem diameter, number of leaves, and image indices (Green Index, Excess Green Index, Excess Red Index, and Color Index of Vegetation Extraction). Results indicated that the triple interaction was significant, revealing the complexity of plant responses to herbicides. Canonical discriminant analysis explained 78.66% of the total variation, with the first canonical discriminant function (29.42%) contrasting structural development and vitality with stress, the second canonical discriminant function (27.59%) reflecting overall plant vigor, and the third canonical discriminant function (21.65%) capturing stress and phytotoxicity negatively affecting growth. The analysis demonstrated that image‐based indices combined with multivariate techniques are effective for quantifying phytotoxicity and distinguishing genotypes (tolerant and sensitive to herbicide effects), overcoming the limitations of visual evaluations, and should be used as a complementary tool to traditional techniques. Therefore, the parental genotype IPR Campos Gerais and the progeny F1A were tolerant to herbicides at different doses, while the parental genotype BAF36 and the progeny F2B were sensitive. Hence, the proposed methodology is effective for identifying herbicide‐tolerant and sensitive genotypes.
Here, we present a single-operator push-cart platform equipped with a 16-beam LiDAR. A push-button interface controls data acquisition, and the data processing pipeline removes ground points, filters noise, performs 5-cm voxelization, and produces plot-level canopy metrics. We validated biomass estimation in hairy vetch (Vicia villosa) and corn (Zea mays) leaf- and whole-plant thinning experiments. In vetch, voxelized estimation of plant volume correlated strongly with destructively measured biomass (r2 = 0.88), showing that the multi-beam LiDAR can produce biomass estimates comparable to previously reported methods. In corn, comparisons of perpendicular (0°) and multi-angle LiDAR beams showed significantly greater voxel counts in the upper canopy when angled beams were used (beam angle × height interaction, p < 0.001), demonstrating that multi-beam scanning provides greater penetration into the upper canopy than a single perpendicular scan plane. We also extended the suite of LiDAR-derived traits to include apparent leaf area index (LAI), mean tilt angle (MTA), persistent homology-based stand density, and plot-bounded foliage area density (FAD). The persistent homology algorithm distinguished between leaf-removal and plant-removal treatments (removal type × removal amount, p = 0.0039). LiDAR-derived LAI has been used to estimate canopy leaf area, but gap-fraction approaches do not fully exploit the ability of LiDAR to resolve distance. Plot-bounded FAD used ray length and interception distance within defined plot volumes and was more sensitive to plot-level treatments than apparent LAI or MTA, detecting differences associated with both the removal amount and removal type. These results show that a robust, portable, multi-beam LiDAR cart can reproduce plot-level canopy measurements and improve trait especially in research-sized plots.
The selection of genotypes adapted to water stress requires experimental facilities that allow environmental control without compromising physiological and yield relevance. The objective of this study was to design and validate an outdoor phenotyping semi-controlled platform, PlaFe, which comprised sixty-two high-volume prismatic lysimeters arranged in rows 1.2 m long and spaced 0.6 m apart. Soil water dynamics were monitored weekly using a weighting system. To validate PlaFe, two soybean genotypes were exposed to two water scenarios for forty days from R2 + 7d, during two growing seasons. Two irrigation treatments were applied: irrigation to keep soil water content over 60–70 % of field capacity (EH0), and irrigation equivalent to 35 % of that applied in EH0 (EH1). Water consumption, crop biomass, and pod number were determined at maturity. On average, water stress reduced both biomass and pod numbers by 40 %. However, reproductive efficiency varied among genotypes. Canopy temperature increased by 0.56 °C as daily water consumption decreased, demonstrating its potential to assess drought. These results demonstrate PlaFe’s potential for the accurate evaluation of crop response and adaptation to diverse water scenarios without compromising the complex plant-environment interactions inherent to field conditions.
Context Precision agriculture can benefit from within-field boundary line analysis (BLA) to identify the most limiting factors impacting crop yield. In this study, the BLA was applied at the within-field scale using yield monitor data and key environmental factors, including evapotranspiration (ET), elevation, and the apparent soil electrical conductivity (ECa). Aims We aimed to develop and evaluate a novel BLA method to estimate Yp and quantify Yg at the within-field scale, and to assess the diagnostic potential of freely available proxy variables for identifying spatially variable yield-limiting factors in dryland wheat production. Methods We combined Mahalanobis distance-based filtering for data denoising with 2Dl kernel density estimation (KDE) and percentile thresholding to select high-density, high-yield points that define the upper yield envelope. A generalised additive model (GAM) was then used to produce the boundary line through these selected points to represent the Yp. Key results Results from the two case studies showed that this approach was robust and less sensitive to noise and outliers in fine-scale datasets. Freely available ET and elevation could be proxies to highlight the impact of some limiting factors, such as frost events or waterlogging. The ECa could identify areas where some potential soil-related factors (e.g. lower clay content reducing plant available water capacity) could be the limiting factors. Conclusions While the proxy variables effectively indicated potential limiting factors, ground-truth validation is required to confirm the underlying causal mechanisms. Implications Growers could benefit from the BLA approach to identify local yield constraints, estimate site-specific Yp and Yg, and fine-tune their inputs, leading to more efficient resource use and improved profitability.
This research introduces a multimodal deep learning framework for early detection of plant pathogens to capture pre-symptomatic biochemical changes in plants while simultaneously modeling the environmental drivers of disease development. A hybrid fusion architecture combines 3D convolutional neural networks for spatial-spectral feature extraction from HSI cubes with transformer-primarily based temporal modeling of climate sequences. Cross-modal attention mechanisms dynamically weight discriminative features, which includes chlorophyll degradation bands and humidity thresholds, to permit joint representation learning. The framework achieved 94.5% accuracy in pathogen detection, outperforming unimodal HSI (84.1%) and climate- only (76.5%) baselines by 10-18 percentage points. Moreover, it detected fungal infections 5-7 days before visual symptom onset and had a 12.3% higher F1-rating compared to the current methods. Field simulations showed that precision application resulted in 41% reduction in fungicide use. By connecting proximal sensing with climatic analytics, this research contributes to precision agriculture by providing timely and eco-friendly pest control of diseases. The multimodal fusion framework is introduced to overcome the limitations of unimodal approaches. It integrates the most appropriate data sources, thus allowing the earliest and most accurate detection of plant pathogens.
Abstract Flower-visiting insect populations are declining since the 1990s, especially because of the decrease of floral resources in agricultural settings. Mass flowering crops can help increase resource availability, and plant breeding can be directed towards selecting varieties attracting more flower-visiting insects. This requires the implementation of an automated high-throughput phenotyping tool for assessing the attractiveness of plant genotypes to flower-visiting insects. In this study, ( i ) we present a procedure to take standardized images of sunflower heads with camera traps continuously at day and night in the field; ( ii ) we trained two versions of a deep learning model, named PolliCrop, to automatically detect and identify the three insect classes visiting the most sunflower (non- Bombus bees, bumble bees, lepidopterans); ( iii ) we assessed and validated the ability of PolliCrop to correctly predict the true visitation frequencies of the insect classes on three sunflower genotypes; ( iv ) we presented two statistical approaches to compare the insect visitation frequencies between plant genotypes, one including weather variables, and the other one without. One PolliCrop version yielded satisfying performance to correctly detect the three insect classes. In particular, it correctly predicted the insect visitation frequencies on two sunflower genotypes in a range of ±10%. The other PolliCrop version can be useful in certain contexts of images and objectives. PolliCrop can be extended in the future to other crop species by training PolliCrop on new images captured in these crops. The field experimental design to set up for comparing the attractiveness between genotypes is also discussed.
0. Morphological traits such as floral area and body size are fundamental to ecological research, serving as inputs for studies of pollinator–plant interactions, habitat quality, and biodiversity monitoring. However, accurately measuring these traits from images remains challenging, particularly in complex field conditions where existing tools exhibit reduced accuracy and limited generalizability across taxa. We present EcoMorph, a modular morphological measurement system that leverages the Segment Anything Model 3 (SAM3) to quantify traits across diverse ecological contexts. Unlike task-specific segmentation models requiring domain-specific training data, SAM3’s prompt-based architecture enables segmentation of arbitrary biological structures from natural-language prompts, using the same underlying model across flowers, insects, and other targets without retraining. From the resulting segmentations, EcoMorph extracts three classes of measurement: area, linear dimensions, and object counts. We validated EcoMorph across two ecological scales. At the intermediate scale, EcoMorph-derived floral area agreed closely with manual ImageJ measurements (R 2 = 0.935, n = 74) under simple-background conditions and (R 2 = 0.928, n = 58) under complex-background conditions, with valid predictions for 95% of images. At the fine scale, EcoMorph-derived insect body area was strongly correlated with hand-measured intertegular distance (r = 0.810, n = 349), capturing body-size variation across species from the small Bombus impatiens to the large Xylocopa virginica . Object counts matched manual counts almost exactly for well-separated insects in an insect box (R 2 = 0.9997, n = 12). By combining prompt-based segmentation with modular measurement, EcoMorph enables high-throughput quantification of area, size, and abundance from heterogeneous image sources without taxon-specific training. This generality supports a broad range of ecological applications, including pollinator and plant trait research, biodiversity and abundance monitoring, and allometric biomass estimation.
Reproduction assets foundThe paper's Data and code availability statement provides a public Zenodo deposit containing the validation datasets and code used for the EcoMorph phenotyping measurements (floral area, insect morphometrics, counts), plus a public web deployment of the EcoMorph software itself.Code · publicValidation datasets and code are available here on Zenodo
https://zenodo.org/records/20980236.Open asset ↗Zenodo · 20980236pdf-page:2 lines:1-54Code / dataset availability confirmedEurope PMC · checked 14 Sept 2026
Ustilago maydis is a biotrophic fungus that causes smut disease in maize, leading to tumor formation on aerial parts of the plant. While U. maydis has been a model for plant-fungal interaction studies, no tool has existed to automatically quantify infection symptoms under laboratory conditions for deep learning analysis. To address this, we developed a rotating camera system that captures videos of plants under customized lighting and shutter settings. These videos were used to train machine learning models to distinguish between healthy and infected plants. Two detection approaches have been presented. In the first approach, by employing a naive masking technique and combining classical machine learning classifiers utilizing handcrafted features, the model achieved a reasonable performance, with an Area Under the Curve (AUC) of maximum 0.90 on the Receiver Operating Characteristic in one of the classifiers, showing relatively high sensitivity and specificity. The second approach utilizes pre-trained YOLO11 model for object detection and further classification. The YOLO11-based approach outperforms traditional methods, achieving near-perfect validation accuracy (AUC: 0.99-1.00), demonstrating its superiority for real-time, scalable applications. Our toolset, featuring a cost-efficient and customizable scanning platform with open building-blocks design, provides a valuable resource as a proof-of-concept for unbiased disease symptom detection and scoring, with potential applications in other plant pathology studies. This point enables easy replication and adaptation by other research laboratories which makes the platform robust, scalable and practical beyond our specific application.
Rice (Oryza sativa L.) is a crucial food crop, supplying a significant portion of the global population's caloric intake. With the shift from traditional breeding methods to digital approaches, image analysis is becoming essential for distinguishing between rice cultivars. However, the optimal growth stages for effectively utilizing image analysis to classify rice varieties remain uncertain. This study aimed to evaluate 102 rice cultivars through non-destructive image processing and RGB ratio analysis. Images were captured every two days throughout the growth period, and an RGB ratio formula was developed, excluding background pixels to focus on plant characteristics. Regression analysis identified critical time points for differentiation, with the red (R) channel being most effective at 55 and 75 days post-transplanting, and the green (G) channel at 60 and 80 days. Hierarchical clustering of slopes from piecewise regression categorized the 102 cultivars into three distinct clusters, representing their ecological types. These findings provide a precise and efficient method for classifying rice cultivars, offering breeders key insights into the most effective stages for variety differentiation. By optimizing image analysis techniques, this research enhances the efficiency of rice breeding programs and supports the targeted management of genetic resources for improved trait selection.
Urban forest health may be monitored and supported with the implementation and maintenance of an accurate community tree inventory. The longitudinal recording of a tree’s attributes (i.e. diameter and height) may inform potential inputs and activities related to maintenance, health, and the establishment of a tree protection zone. Urban tree inventories feature barriers to implementation including resources (i.e. labour, time, and finances), competing priorities, and gaps in knowledge. The objectives of this case study were to investigate the feasibility of structure-from-motion (SfM) approaches to measure tree height and trunk diameter (1) with oblique RGB imagery, (2) with and without GCP, and (3) during leaf-off and leaf-on conditions. Structure-from-Motion datasets were obtained using Unoccupied Aerial Systems (UAS) – “drones” and related components – to collect oblique aerial imagery, which was processed with photogrammetry software Agisoft Metashape. Tree height measurements using leaf-on imagery (R2 = 0.58, RMSE = 1.34 m) were more accurate when compared to leaf-off imagery with Ground Control Points (GCP) (R2 = 0.47, RMSE = 3.41 m) and leaf-off imagery without GCPs (R2 = 0.43, RMSE = 3.49 m). Tree height measurements during the leaf-off period had no significant difference when comparing imagery with and without GCPs. Trunk diameter measurements using leaf-off imagery were not significantly different with the use of GCPs (R2 = 0.68, RMSE = 6.39 cm) compared to those without (R2 = 0.68, RMSE = 7.85 cm). This case study highlights the applicability and accuracy of Unoccupied Aerial Systems and Structure-from-Motion methods when collecting important urban tree inventory parameters, and presents an accessible, reliable, and replicable workflow for urban forestry practitioners with limited photogrammetry-related experience.
Abstract Precision field management and high-throughput plant phenotyping increasingly rely on remote sensing to capture spatial and temporal variability in crop performance. Unmanned aerial vehicle (UAV) – based sensing offers unique advantages for field-scale data collection, including high spatial resolution, flexible deployment, and scalable throughput. However, the full potential of UAV platforms remains constrained by labor-intensive operations across flight execution, data transfer, and processing workflows. This study presents a systematic evaluation of an automatic UAV-based crop sensing platform through a season-long, multi-crop field experiment. Data acquisition was conducted over a maize irrigation trial and a soybean breeding experiment, resulting in 176 completed flights over 28 days during the growing season. High-frequency flights on selected days captured diurnal dynamics in key canopy traits, including maize leaf rolling under drought stress and genotype-dependent plot temperature variation in soybean. In the maize irrigation experiment, significant differences in diurnal canopy cover ratio (CCR) were observed among irrigation treatments. The predictive relationship between CCR and final grain yield strengthened throughout the day, with the coefficient of determination (R 2 ) increasing from 0.05 in the early morning (RMSE = 3.05 Mg ha − 1 ) to 0.65 at midday (RMSE = 1.87 Mg ha − 1 ), highlighting the importance of temporal optimization in UAV-based sensing. Temperature measurements from the onboard thermal infrared camera showed a strong overall linear correlation with ground truth measurements (R 2 = 0.85). In the soybean trial, the highest plot temperature was observed on the fast-wilting genotype. Additionally, regression models were developed to estimate key crop traits, including canopy height (CH) and leaf area index (LAI), demonstrating the platform’s quantitative sensing capability. Overall, this study demonstrates that automatic UAV systems enable high-temporal-resolution crop monitoring while substantially reducing operational cost. The results highlight their potential for precise crop management and scalable field phenotyping. Future work will focus on integrating automated data processing pipelines to support near-real-time analytics and decision-making.
Abstract Multispectral three‐dimensional (3D) imaging offers substantial potential for plant phenotyping, yet manual segmentation of plant organs remains a bottleneck in breeding programs. We developed a color‐based filtering workflow for faba bean ( Vicia faba L.) point clouds that optimizes lower and upper thresholds of spectral indices and broadband brightness via Bayesian optimization. Rather than maximizing geometric segmentation accuracy, thresholds are selected to maximize correlations between retained points and yield‐related traits, outperforming manual filtering, reducing user effort, and standardizing decisions. Across multispectral 3D point clouds, Bayesian optimization recovered index‐specific threshold ranges that yielded strong in‐sample correlations with grain yield ( r = 0.72), bean number ( r = 0.61), pod number ( r = 0.53), and straw biomass ( r = 0.72). Peak associations occurred at harvest for straw biomass, at 41 days before harvest (DBH) for seed yield, 33 DBH for bean number, and 34 DBH for pod number. Across the season, greenness‐based indices and broadband brightness metrics consistently showed stronger links with seed yield than pigment ratio or water status indices. For straw biomass and pod number, pigment ratio indices showed consistently lower correlations. Targeting trait‐relevant canopy signals via Bayesian optimization enables reliable, nondestructive assessment of relationships between spectral signals and yield‐related traits in faba bean. By optimizing thresholds to maximize trait correlations rather than geometric accuracy, the workflow can support earlier, more cost‐efficient identification of high‐performing genotypes under drought stress and contribute to strengthening high‐throughput phenotyping in breeding programs. This enables faster identification of canopy signals most relevant to target traits.
ABSTRACT While whole-genome sequencing captures millions of single nucleotide polymorphisms (SNPs) and hyperspectral imaging (HSI) enables non-destructive plant phenotyping, integrating these modalities to link genotype to phenotype remains challenging due to their high dimensionality and non-linearity. This study presents DeepPheno a deep learning framework that predicts SNP genotypes from HSI data, using model predictability as a proxy for genotype-phenotype association. HSI data were acquired from 194 lettuce genotypes under field conditions. HSI data patches (20×20 pixels × 224 spectral bands) were used to train a hybrid CNN to predict the variant of a specific SNP. The framework was validated on SNPs with known phenotypic effects (anthocyanin, leaf serration, pale pigmentation), achieving high predictive performance (AUC ranging from 0.806 to 0.935), whereas models trained on randomly shuffled labels performed at chance (mean AUC ≈ 0.51). Extending the workflow to 50 randomly selected putatively neutral SNPs, most yielded low predictability, but two showed high performance (AUC > 0.76), suggesting uncharacterized genotype-phenotype links. Explainable AI, including SHAP and Grad-CAM, identified relevant spectral and spatial features driving these predictions, particularly the green and red-edge wavelengths associated with pigment dynamics and leaf structure. These results establish a framework for understanding complex genotype-phenotype interactions in plants and extracting these links from HSI data without predefining the exact trait values. It provides an avenue for high-throughput trait discovery and description and extends the integration of image-based phenomics with plant genetics.
Reproduction assets foundThe paper's Data Availability statement deposits authors' code, scripts, and supplementary material in a public GitHub repository, including a downscaled de-identified sample dataset demonstrating the pipeline. The raw HSI/genotype datasets are proprietary under NDA and not public.Code · publicThe code, scripts, and supplementary material supporting the findings of this study have been deposited in the GitHub repository at https://github.com/frankgyan/Utrecht-University--HSI .Open asset ↗frankgyan/Utrecht-University--HSIlines:195-223Code / dataset availability confirmedOpenAlex · Europe PMC · Crossref · checked 5 Sept 2026
Sweet potato virus disease (SPVD) is one of the most destructive diseases affecting sweet potato production worldwide, causing severe yield losses and posing a significant threat to food security. Vision-based intelligent diagnosis has emerged as a promising solution for large-scale SPVD monitoring due to its low cost and scalability. However, existing publicly available datasets for SPVD are extremely limited and typically focus on a single task, such as disease classification or lesion segmentation, under constrained imaging conditions. This lack of comprehensive, task-oriented datasets significantly restricts the development, evaluation, and fair comparison of advanced computer vision methods for SPVD analysis. In this study, we present SPVD-Field, a task-oriented multi-task visual dataset suite composed of two independently collected sub-datasets optimized for different computer vision tasks. Rather than constructing a single homogeneous dataset, SPVD-Field is deliberately organized into two complementary task-oriented sub-datasets: SPVD-DET, designed for disease detection with bounding-box annotations, and SPVD-SEG, designed for fine-grained lesion segmentation with pixel-level masks. The two sub-datasets were independently collected using different acquisition protocols optimized for their respective tasks, while sharing a unified semantic definition of SPVD symptoms, crop growth stages, and field environments. SPVD-Field captures substantial real-world variability in imaging scale, viewpoint, illumination, background complexity, and symptom manifestation, reflecting the inherent challenges of fieldbased disease diagnosis. We provide detailed documentation of data acquisition, annotation strategies, and quality control procedures, along with baseline benchmark results for both detection and segmentation tasks to demonstrate the usability and difficulty of the dataset. By offering a structured dataset suite rather than a single-task collection, SPVD-Field aims to support diverse research directions, including detection, segmentation, multi-task learning, and disease severity analysis, and to facilitate reproducible and comparable research in SPVD-related plant phenotyping.
Reproduction assets foundThe paper's core asset is the SPVD-Field dataset (SPVD-DET detection images with bounding-box annotations and SPVD-SEG segmentation images with pixel-level masks), explicitly deposited in a public repository via the data availability statement with a DOI link.Dataset · publicThe datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found below: https://dx.doi.org/10.21227/hq1q-jp43 .Open asset ↗10.21227/hq1q-jp43lines:664-703Code / dataset availability confirmedCrossref · checked 15 Sept 2026
Accurate estimation of evapotranspiration (ET) is critical for irrigation management in water-scarce regions such as the Middle East and North Africa (MENA). This study compares sensible heat flux (H), latent heat flux (LE), and ET derived from eddy covariance (EC) and a boundary-layer scintillometer (BLS) operated with an aperture reducer, deployed simultaneously over an irrigated late-season potato field (1.8 ha) in the Beqaa Valley, Lebanon. Satellite NDVI observations indicate that the BLS–EC overlap period (13 October–27 November 2021) sampled the crop from peak canopy (NDVI ≈ 0.85–0.90) through the onset of senescence (NDVI ≈ 0.79). The BLS (Scintec BLS900) operated along a 140 m path. The EC system showed incomplete daytime energy-balance closure, with a regression slope of ≈0.69 and a seasonal Bowen-ratio-preserving correction factor of CF = 1.24 (a ~19% closure deficit) was used. Across the matched period, daily H from the BLS was strongly correlated with EC (r ≈ 0.82) but systematically lower, with a regression slope of ≈0.63 that persisted across timescales; this scale-invariant amplitude compression reflects the path-averaged, similarity-based nature of the scintillometer retrieval rather than the EC closure deficit, which instead governs the mean bias. BLS-derived daily ET showed a systematic positive bias relative to uncorrected EC (mean bias error, MBE = +0.30 mm d−1; +16% cumulative). Applying the Bowen-ratio-preserving correction (CF = 1.24) to EC reduced this to MBE = −0.14 mm d−1 (−6%), and the residual-to-LE correction yielded MBE = −0.15 mm d−1 (−6.4%); the latter comparison is only partly independent, as both methods share the same Rn and G. The Bowen-ratio-preserving method is therefore recommended for this dataset. Overall, the BLS captured the temporal variability of crop water use well, but residual-based ET estimates require careful treatment of the energy-balance-closure gap and are sensitive to the high BLS gap fraction (61.6% of 15 min records over the overlap, exceeding 90% at night). Once EC is closure-corrected to serve as the reference, the BLS offers a cost-effective alternative for field-scale ET monitoring in the MENA region, subject to the conditional agreement documented here.
Reproduction assets foundThe paper's flux/ET datasets are only available on request from the corresponding author, so they do not qualify as public assets. However, the Supplementary Information file (available at the MDPI supplementary URL) explicitly contains experiment sensor documentation and field/canopy images (Figures S1–S4: study site,Supplement · publicmeasurements along the beam. Because these results derive from a single crop, season, and phenological window, their generalization awaits multi-site, multi-season replication spanning the full-canopy cycle—the priority for subsequent campaigns.
Supplementary Materials
The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/s26144398/s1 , Figure S1: Study site and potato canopy—Beqaa Valley, Lebanon; Figure S2: Eddy covariance system—full tower view (peak canopy); Figure S3: EC sensor suite close-up and soil sensor installation; Figure S4: BLS900 scintillometer—transmitter, receiver, and meteorological station.
Author Contributions
Conceptualization, HOpen asset ↗lines:251-268Plant phenotyping relevance matchCrossref · checked 8 Sept 2026
Marcel M. El Hajj · Kasper Johansen · Oliver M. Lopez Valencia · Fabio Veiga de Camargo · Yu-Hsuan Tu · Samer K. Al Mashaharawi · Omar A. López Camargo · Victor Angulo · Dominique Courault · Matthew F. McCabe
Purpose In recent years, there has been a growing use of unmanned aerial vehicle (UAV) based light detection and ranging (LiDAR) data for mapping plant area index (PAI) in orchards. However, using LiDAR time-series collected throughout the growing season to assess PAI variations in response to phenology, represents an understudied area of investigation. Furthermore, establishing the optimal spatial resolution for mapping biophysical variables of tree crops from LiDAR point cloud data remains poorly defined. Here, we assess the capability of a UAV-based LiDAR system to characterize cherry trees throughout the growing season, with a focus on monitoring PAI and the vertical structure of individual trees. Methods A time-series of 14 point cloud acquisitions with a density of 3300 points/m2 was collected between February and December 2022, covering all phenological stages of a cherry orchard in southern France. A voxel-based method was applied to create a three-dimensional grid within which PAI was estimated for each voxel. PAI was mapped by accumulating the individual voxel-based PAI values within each vertical voxel column. Results The results demonstrate that a voxel size of at least 0.7 m is required to retrieve reliable PAI estimates (RMSE = 0.58 m2.m−2, MAE = 0.48 m2.m−2, bias = 0.19 m2.m−2, rRMSE = 23%, and R2 = 0.51), while a voxel size of 1 m produced the most accurate PAI estimates (RMSE = 0.5 m2.m−2, MAE = 0.41 m2.m−2, bias = 0.07 m2.m−2, R2 = 0.59), when assessed against field-based PAI measurements obtained with a LAI-2200 Plant Canopy Analyzer. The temporal variation of canopy PAI illustrated the progression of key phenological stages, including flowering, leaf development, ripening and senescence, as well as the response of the canopy to drought stress (reduction in PAI due to leaf rolling) during the summer. The maps of PAI successfully described the variations in leaf canopy density for different cherry varieties and allowed assessment of the vertical PAI profile at the individual tree level, which provides valuable insight into tree condition. Conclusion This study confirms that seasonal UAV-LiDAR monitoring is a viable, informative approach for capturing orchard canopy dynamics at the individual tree and sub-canopy level, linking canopy structure to phenology, varietal differences, and stress responses across the growing season.
Abstract Plant diseases affecting leaves and fruits cause substantial yield and economic losses worldwide, particularly in horticultural crops cultivated under diverse agro-climatic conditions. Early and accurate disease diagnosis is essential for effective crop management; however, manual inspection is time-consuming, subjective, and often infeasible at large scale. In this work, we present the Tomato–Chilli–Papaya (TCP) Fruit and Leaf Disease Dataset, a comprehensive multi-crop image dataset designed to support deep learning-based plant disease recognition. The dataset comprises labeled RGB images of healthy and diseased leaves and fruits from three economically important crops—tomato, chilli, and papaya—captured under real-field and semi-controlled environments, reflecting significant variability in illumination, background complexity, and disease severity. To demonstrate the applicability of the dataset, several commonly used convolutional neural network (CNN) architectures, including VGG, ResNet, DenseNet, MobileNet, and EfficientNet models, were trained and evaluated on the TCP dataset using transfer learning. Experimental results show that deep CNN models can effectively learn discriminative visual features corresponding to disease-specific patterns such as leaf spots, lesions, discoloration, curling, and fruit surface abnormalities. Lightweight models such as MobileNet achieve competitive performance with reduced computational cost, while deeper architectures provide improved accuracy at the expense of higher complexity. The results highlight the importance of dataset diversity for robust model generalization across multiple crops and plant organs. The TCP dataset provides a challenging benchmark for single-crop and multi-crop disease classification and supports the development of advanced deep learning, attention-based, and explainable AI models for precision agriculture. By enabling reproducible research and realistic performance evaluation, this dataset contributes toward scalable and practical AI-driven plant disease diagnosis systems aimed at reducing yield losses and supporting sustainable agriculture.
Reproduction assets foundThe paper introduces the TCP (Tomato-Chilli-Papaya) fruit and leaf disease image dataset and reports CNN experiments on it. The dataset is publicly deposited on Mendeley Data, and the authors state that analysis code is available on GitHub. Both are paper-specific, public, and actionable.Dataset · publicData is available on Mendeley:1Open asset ↗pdf-page:27 lines:1-51Code · publicCode availability: Code is available on GitHub 2Open asset ↗GitHubpdf-page:27 lines:1-51Plant phenotyping relevance matchOpenAlex · Crossref · Europe PMC · checked 15 Sept 2026
Sergey Shabala · Ping Yun · Zhong‐Hua Chen · M M Zhou · Min Yu · Nadia Bazihizina
Field / plotCell / cellular structureWhole plant / canopy / plot / fieldGrowth / development / phenologyStress response / tolerance
Abiotic stress tolerance has been significantly weakened in modern crops during the domestication process. Regaining tolerance has become a critical task in light of current climate trends and their impact on global food security. Abiotic stress tolerance is an extremely complex trait and is conferred at various levels of plant functional organization and developmental stages, with regulatory mechanisms operating across multiple scales, from individual cells to tissues and the entire plant. The emergence of advanced molecular tools such as single-cell RNA sequencing and spatial omics technologies has revolutionized the field, advancing our understanding of plant responses to hostile environments. However, the implementation of this knowledge in crop breeding programmes is handicapped by the lack of appropriate phenotyping platforms. Here, we argue that current phenotyping methods may be excellent tools for functional validation of previously discovered traits but have limited predictive value in stress biology. We also propose that bridging the mismatch between omics technologies and phenotyping is the only way to account for cell-specific operation of key genes conferring stress tolerance and implementing them in breeding programmes. Some practical examples using cell-based phenotyping tools such as fluorescence dyes or electrophysiological methods are given, and current limitations and prospects of cell-based phenotyping are discussed.
Introduction Canopy water content (CWC) is an important indicator of crop water status **and** supports precision irrigation decision-making. Plot-level CWC estimation using UAV imagery often relies on canopy mean features, whereas the role of within-plot canopy-signal distributional information remains insufficiently examined. Methods In this study, spring maize at the Shiyanghe site was monitored using UAV-based multispectral and thermal infrared imagery. Mean, percentile, and dispersion features were extracted from effective canopy pixels within each plot. RFECV feature selection, 50 repeated random train-test splits, paired statistical tests, simulated spatial aggregation, and four regression models were used to evaluate the stage- and scale-dependent contribution of these features. Results and discussion Water stress affected both overall spectral-thermal responses and within-plot signal distributions. Before tasseling, percentile and dispersion features were frequently selected and provided complementary information, especially for tree-based models and finer aggregation scales. After tasseling, mean features generally showed more stable performance, although some distributional features still contained CWC-related information. The supplementary Xinxiang site-internal analysis suggested that, under weak water-gradient and small-sample conditions, distributional features may be frequently selected but may not consistently improve prediction accuracy. Overall, the contribution of distributional features was growth-stage-, scale-, and model-dependent.
Tree growth determines how much CO2 is sequestered from the atmosphere and temporarily stored in woody biomass. At the same time tree growth is affected by increasing temperatures, more frequent drought periods, late frosts and other extreme events associated with climate change. While continuous measurements of radial (secondary) tree growth using dendrometers are well established, monitoring of shoot elongation (primary growth) has largely been neglected because suitable measurement techniques are lacking. As a result, the effects of climate change on primary tree growth remain insufficiently understood. This work aims at reconstructing native deciduous trees in 3D as a basis for measuring and monitoring shoot elongation over entire tree canopies. Here we explored the use of low-cost UAV photogrammetry and of a multi-camera CraneCam system under real-world conditions. Data were collected in two study areas over an entire growing season. We present sensor evaluations, photogrammetric data acquisition and processing strategies. A special focus is placed on the analysis of the resulting photogrammetric 3D point clouds in terms of accuracy, resolution and completeness. Results demonstrate 3D point accuracies of 5-6 mm for entire trees using consumer-grade UAVs weighing less than 250 g and a 3D reconstruction completeness between 92% and 98% depending on the UAV type. The paper introduces a novel 3Dprinted ground-truth branch to evaluate the capability to reconstructing fine-detail structures such as thin tree shoots. Finally, we discuss operational challenges and initial experiments towards a skeletonization of entire trees based on photogrammetric point clouds.
A bstract Quantifying root traits such as root length (RL) and root surface area (RSA) from minirhizotron imagery is a valuable approach for overcoming the phenotyping bottleneck that limits understanding and improvement of crop productivity, resource use efficiency and resilience in field experiments. However, current approaches remain labor-intensive, and deep learning (DL) methods suffer from limited generalization ability. We present RootQuant, an end-to-end DL model that simultaneously predicts RL and RSA directly from minirhizotron images using only whole-image trait values as supervision, thereby eliminating the need for pixel-level annotations. The model’s generalization ability was evaluated across species and fine-tuning configurations. The practical applicability of the model was further assessed under field conditions by converting image-derived RL estimates into volumetric root length density (vRLD). Using 118,191 maize and soybean images collected between 2009 and 2020, RootQuant trained on both species achieved an R 2 of 0.90 and an RMSE of 2.9 mm for RL, and an R 2 of 0.88 and an RMSE of 4.2 mm 2 for RSA. The same mixed-species model generalized strongly across species, yielding an 8% relative improvement in R 2 and a 30% lower RMSE on maize compared with the same architecture trained on a single species and applied zero-shot. Image-derived RL predictions converted to vRLD showed the expected depth-dependent decline in vRLD, as was also found by coincident destructive quantification of roots washed out of soil cores. By providing a generalist backbone model trained on a large dataset from two major crop species, RootQuant enables high-throughput simultaneous estimation of two relevant root traits directly from raw imagery without task-specific fine-tuning, thereby accelerating in situ root system analysis and phenotyping applications.
Introduction: Plant phenotyping requires accurate and repeatable three-dimensional structural information, but practical acquisition conditions in greenhouses, seedling rooms, and indoor pot experiments often include complex backgrounds, handheld motion blur, and thin leaf structures. These factors reduce the robustness of conventional three-dimensional reconstruction methods and limit their use in low-cost and automated phenotyping. Methods: To address this problem, this paper proposes F2DMAS, an automated three-dimensional plant phenotyping workflow using consumer-grade smartphone videos. The workflow first converts multiview RGB videos into image sequences and removes motion-blurred frames through frequency-domain quality filtering. A frequency-spatial plant segmentation module, termed FSAM3, is then introduced to separate plant structures from complex backgrounds without task-specific annotated training data. The segmented image sequences are further reconstructed using 2D Gaussian Splatting, followed by TSDF-based meshing, scale recovery, and virtual measurement for extracting plant height, canopy width, leaf length, and leaf width. Results: Experiments were conducted on 15 plant species under two acquisition scenarios. The proposed workflow achieved stable plant reconstruction under non-ideal background conditions, with PSNR, SSIM, and LPIPS values of 31.09, 0.9711, and 0.0365, respectively. Compared with the baseline reconstruction workflow, F2DMAS substantially reduced the processing time for mesh extraction while improving reconstruction quality. The extracted phenotypic traits showed strong agreement with manual measurements, with R² values ranging from 0.90 to 0.99, RMSE values ranging from 0.64 to 1.21 cm, and MAPE values ranging from 4.50% to 9.73%. Discussion: These results indicate that F2DMAS can provide an end-to-end workflow from smartphone video acquisition and plant segmentation to three-dimensional reconstruction and phenotypic trait extraction. The proposed method offers a practical and deployable solution for greenhouse seedling cultivation, potted plant experiments, and low-cost three-dimensional plant phenotyping.
Accurate lesion segmentation is essential for automated plant disease analysis in precision agriculture. Although the Segment Anything Model (SAM) exhibits strong generalization ability, its direct application to plant disease images in natural field environments remains challenging due to cluttered backgrounds, dense leaf veins, uneven illumination, and frequent occlusions. In particular, SAM mainly relies on global structural cues and is often insufficiently sensitive to subtle lesion textures and weak local details, which can result in missed small or early-stage lesions and inaccurate boundary delineation. To address these limitations, we enhance SAM with a disease-specific detail compensation module for plant disease lesion segmentation. A ResNet-50based branch is employed to extract fine-grained local texture features that are difficult for SAM to capture. These fine-grained features are fused with SAM encoder representations and then injected into the SAM decoder, enabling more accurate lesion prediction while preserving SAM’s strong global modeling capability. More importantly, we propose a reliability-guided variational fusion framework to further improve the interaction between heterogeneous features. Specifically, instead of conventional similarity or addition-based fusion, we introduce an uncertainty-aware variational fusion strategy that explicitly quantifies the confidence of each feature stream. An uncertainty encoder models feature distributions probabilistically, and a variational fusion module dynamically assigns higher weights to more reliable features while suppressing uncertain or interfering responses. In addition, Kullback-Leibler divergence regularization is introduced to stabilize cross-feature alignment and improve fusion robustness. Extensive experiments on PlantSeg, PlantDoc-Seg, and ATLDSD demonstrate that the proposed method outperforms state-of-theart approaches, achieving DSC scores of 81.05%, 91.12%, and 88.27%, respectively. The proposed method addresses SAM’s weakness in fine-grained disease feature extraction, accurately identifies early and small lesions, and delivers reliable segmentation for field plant disease automatic diagnosis.
Reproduction assets foundThe paper evaluates ReLeaf-SAM on three public plant disease segmentation datasets. One of them, PlantDoc-Seg, is explicitly a community-provided Kaggle dataset with a verbatim URL matching an allowed URL; it is a public plant image/mask dataset directly used for this paper's segmentation measurements. PlantSeg and ATLDataset · publicTherefore, we used a community-provided segmentation subset from Kaggle 1 , which we refer to as PlantDoc-Seg in this study. This subset is derived from PlantDoc and contains 588 diseased leaf images with corresponding binary masks, enabling supervised leaf disease segmentation.Open asset ↗Kagglelines:48-115