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.
Junjie Wu · Bei Zhou · Jie Liu · Zhuyang Xie · Jie Luo · Jiajun Liu · jiahui zhu · Mingju Li · Yan Ai · Mingwei Liu · Yu Chen
MaizeRapeseed / canolaSoybeanLiDAR / point cloudLeafStem / branch
Three-dimensional plant point clouds retain crop architecture, but organ-level phenotyping requires reliable semantic separation of leaves and stems. Plant-GeoAT is a geometry-aware parsing network that encodes local relative-XYZ neighbourhoods before RGB fusion and couples spatial neighbourhoods with feature–space relations for dense point prediction. We evaluated the model separately within the native protocol of a self-built structure-from-motion rapeseed dataset, an image-based soybean dataset, and laser-scanned Pheno4D maize and tomato datasets; these are within-dataset train/test experiments, not cross-dataset transfer or domain-generalisation tests. Across five seeds, mIoU was 92.26 ± 0.28%, 82.50 ± 0.24%, 99.74 ± 0.05%, and 94.75 ± 0.15%, respectively. Maize Stem IoU reached 99.57 ± 0.09%. Adding LLGE increased the four-dataset average mIoU from 66.38% to 85.84%, and the complete LLGE + SSCA model reached 92.31%. On the fixed six-sample test set, exploratory semantic-guided clustering achieved 86.67 ± 7.45% Count Accuracy; structural correctness ranged from 3/6 to 4/6 samples across seeds. Plant-height consistency was assessed independently on all 60 reconstructed rapeseed samples. The results indicate that geometry-to-context encoding produces organ-level semantic units for subsequent phenotyping, while independent instance-labelled datasets and additional growth stages are still needed for trait-level validation.
Reproduction assets foundThe paper's Plant-GeoAT research-preview code (model implementation, training and evaluation scripts) is publicly available at an immutable GitHub commit under the authors' kimevans111 account. The self-built rapeseed point-cloud dataset and annotations are not public (available only on request), and third-party publicCode · publictheir original providers and publications; this article does not redistribute
third-party datasets. The processed summary metrics required to evaluate the reported results are
included in the main article and Supplementary Material. The reviewed Plant-GeoAT research-
preview code is available at the immutable historical commit https://github.com/kimevans111
/Plant-GeoAT/commit/e0aba5b94e26ba4306d0f7f9901185883fe8d4c4 (accessed on 30 August 2026).
This snapshot is distinct from the repository’s current default branch. It contains the model implemen-
tation, training and evaluation scripts, data-format and release documentation, command templates,
citation metadata, and licence/notiOpen asset ↗Plant-GeoATpdf-raw-page:24 lines:1-50Plant phenotyping relevance matchEurope PMC · checked 17 Sept 2026
Cell division patterns shape developing organs, but their systematic analysis remains limited by the technical difficulty of existing detection methods. Here, we present a practical method for inferring recent daughter-cell pairs from static cell shapes. Although cell divisions have long been inferred by expert judgment, the accuracy of this approach has rarely been quantified. Using live imaging of Arabidopsis leaf primordia, we evaluated 49 features from 7,573 neighboring cell pairs, including 1,114 direct daughter pairs and 6,459 non-daughter pairs. A simple junction-angle criterion correctly inferred 1,030 daughter pairs with 94.1% precision and 92.5% recall, whereas symmetry-based indices performed poorly. Temporal analysis revealed that junction angles were initially high after division, decreased during growth, and were altered by subsequent divisions, thus, separating daughter pairs from non-daughter pairs. This method also achieved 98.4% precision and 97.2% recall in an independent shoot apical meristem live-imaging dataset. Application to the pointed-tip Arabidopsis mutant rpl4d-3 and two non-model species further supported its potential applicability for analyzing division patterns in diverse leaf morphologies. For community use, we provide FIJI/ImageJ and Qt-based graphical user interface implementations. Overall, this study quantitatively validates empirical geometric cues and provides a practical framework for inferring daughter pairs from static tissue images.
Premise Herbarium specimens are increasingly used to extract morphological traits for ecological and evolutionary studies, yet the effects of tissue desiccation on trait measurements remain poorly understood. Here, we tested two hypotheses: (H1) whether higher tissue water content leads to greater measurement changes after herborization and (H2) whether fresh trait values can be reliably predicted from measurements of herborized specimens. Methods We evaluated the reliability of herbarium-based measurements by comparing fresh and dried traits of leaves, flowers, fleshy fruits, and seeds across 262 individuals representing 133 Neotropical Myrtaceae species. Phylogenetic least squares models and machine learning regressions were used to test H1 and H2. Results Leaves and flowers generally shrank after herborization, fruits size metrics tended to increase, and seeds were largely unaffected. Water content was significantly associated with the magnitude of herborization effects in flowers and some leaf and seed traits. Fresh trait values were accurately predicted from measurements of herborized specimens. Prediction errors were lowest for leaf traits, followed by fruits, flowers, and seeds. Discussion These results partially support H1 and support H2, indicating that herborized specimens can be reliably used for trait analyses when organ-specific responses are considered, providing a practical framework to account for potential desiccation bias in functional trait research.
Reproduction assets foundThe paper's trait measurements (Appendix S6 species mean data) and analysis code are openly available in the authors' GitHub repository, explicitly stated in the Data Availability Statement.Code · publice also acknowledge the Sociedade Botânica do Brasil (SBB) for financial support through the Scientia Amabilis 2024 grant.
DATA AVAILABILITY STATEMENT
The data supporting the findings of this study are available in the Supporting Information of this article (Appendix S6 ). Additionally, all data and code are openly accessible at https://github.com/ykilsztajn/fresh_dry_myrtaceae .
REFERENCES
Bewley , J. D.
,
K. J.
Bradford
,
H. W.
Hilhorst
, and
H.
Nonogaki
. 2012 .
Development and maturation
. In
Seeds: Physiology of development, germination and dormancy , 3rd ed. , 27 – 83 .
Springer
,
New York, New York, USA
.
Blonder , B.
,
V.
Buzzard
,
I.
Simova
,
L.
Sloat
,
B.
Boyle
,
R.
Lipson
,
B.
AguiOpen asset ↗ykilsztajn/fresh_dry_myrtaceaelines:268-513Plant phenotyping relevance matchCrossref · checked 17 Sept 2026
Published15 Sept 2026Black Sea Journal of Agriculture
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.
Traditional weight measurement methods need cutting or weighing of the fruit, and this is not practical for preharvest evaluation and market transactions. The suggested approach gives a practical solution for farmers and sellers by giving accurate weight predictions using only external characteristics. Modern technologies, especially artificial intelligence, data analytics and machine learning, are making big changes in the agricultural area. One of the machine learning models used in agriculture is Artificial Neural Networks (ANN). ANN became an important tool in analyzing agricultural data, predicting plant growth processes, finding diseases, determining the effects of environmental factors, reaching productivity goals and detecting weeds and harmful plants. A dataset is created by comprehensively examining 52 watermelons. The dataset includes acoustic properties, geometric measurements, and visual characteristics. The dataset is divided into 40 training samples and 12 test samples. Balanced representation is ensured by using stratified sampling when selecting test samples. The Min-Max normalization method is used for data preprocessing. Nine different training algorithms are comprehensively evaluated within the scope of the study. Eight critical parameters of the ANN models are optimized using four different optimization algorithms to increase the accuracy rate and avoid overfitting. Artificial Bee Colony (ABC), Artificial Fish Swarm Algorithm (AFSA), Whale Optimization Algorithm (WOA) and Grey Wolf Optimizer (GWO) are used as optimization methods. Assessed by five-fold cross-validation, the best configuration (One Step Secant with WOA) achieved a mean absolute error of 0.92 ± 0.28 kg and an RMSE of 1.23 ± 0.37 kg, corresponding to about 10% of the mean fruit weight. Developing a real-time mobile application using the optimized best model will provide practicality in large-scale agricultural enterprises, controlled environments such as greenhouses, and agricultural markets.
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.
Nadeem Fareed · Carlos Alberto Silva · Alexander J. Gaskins · Susan J. Prichard · Andrew T. Hudak · Jinyi Xia · Cesar Ivan Alvites Alvites Diaz
Field / plotLiDAR / point cloudLeafStem / branchMorphology / geometry measurementSegmentationArchitecture / morphology / geometry
Terrestrial Light Detection and Ranging (LiDAR) provides detailed three-dimensional (3D) observations of forest structure, yet transferable wood–foliar semantic segmentation remains challenging because of forest structural heterogeneity, occlusion, and variability across forest ecosystems and terrestrial LiDAR platforms. Existing Deep Learning (DL) approaches are commonly developed for localized forest conditions or complex multi-class semantic taxonomies, often limiting their transferability to structurally distinct forests. Here, we introduce points2SBL, a geometry-driven, terrestrial LiDAR platform-agnostic DL framework that reduces heterogeneous forest vegetation to a transferable wood–foliar representation and subsequently uses these semantics as the foundation for 3D forest fuel characterization. The framework was developed using globally harmonized benchmark datasets acquired across multiple terrestrial LiDAR systems and broadleaf, coniferous, mixed, regenerating, and structurally complex forests. Three point-based architectures (PointNet++, PointNeXt, and Point Transformer) were benchmarked under identical training and evaluation protocols. Point Transformer provided the most consistent cross-dataset performance, achieving overall accuracy (OA) of 92–94%, mean Intersection-over-Union (mIoU) of 0.78–0.87, macro F1-score of 0.87–0.93, and Matthews correlation coefficient (MCC) of 0.74–0.86 across five independent benchmark datasets. Species-level, vertical-profile, and qualitative cross-ecosystem assessments further demonstrated consistent preservation of wood–foliar semantics throughout the 3D forest fuel continuum, including previously unseen forest point clouds, while revealing that some apparent semantic disagreements originated from manual annotation of benchmarks. Finally, the points2SBL wood–foliar semantics were geometrically decomposed using a geometric algorithm named “Points2woodyseg” into fuel-related components: stem, branches, leaf, surface wood, and surface foliage. Collectively, these results demonstrate that a simplified wood–foliar abstraction can provide a transferable semantic foundation for characterizing structurally heterogeneous terrestrial LiDAR point clouds and deriving essential 3D forest fuel information across diverse forest ecosystems.
Diaz-Garcia, L. · Munoz, J. R. · Torres-Lomas, E. · McElrone, A. J. · Sharma, S. · Lupo, Y.
GrapevineRootMorphology / geometry measurementSegmentationGrowth / time-series analysisGrowth / development / phenologyRoot system architectureWater status / transpiration
Root system architecture shapes how grapevine rootstocks take up water and nutrients, yet roots remain the least phenotyped grapevine organ because they are hidden and hard to image. We present a low-cost phenotyping pipeline that pairs custom acrylic rhizotrons (about US$30 each) with a consumer flatbed scanner and BiRefNet, a general-purpose deep-learning model used without training on root images, followed by automated mask cleaning, skeleton-based trait extraction, and soil moisture mapping. We tested it on nine commercial rootstocks scanned 16 times over 42 days after transplanting (DAT), with half under a ten-day water deficit. From 1,108 images we extracted 21 whole-root, depth-resolved, and topological traits. Genotypes differed in nearly every trait and in how they changed over time. Heritability of size and branching traits peaked at 0.92-0.93 between 21 and 31 DAT and fell for width, depth, and convex hull once roots reached the rhizotron walls, defining the best measurement window. The image-derived soil moisture map accurately tracked the deficit and its recovery. Deficit plants shifted new root growth to deeper soil without growing less overall, and the substrate dried fastest around older and denser roots. Root brightness decreased with root age and local moisture, and transport segments (axes serving several tips) were brighter than terminal laterals in every genotype. Root system size was associated with stomatal conductance in well-watered plants, and stomatal recovery after re-watering correlated with new root growth. The pipeline turns simple hardware into a quantitative, time-resolved root phenotyping platform suitable for breeding.
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.
Automated wheat-seed detection and counting play important roles in high-throughput plant phenotyping, seed characterization, and agricultural research. Conventional manual counting is labor-intensive, time-consuming, and susceptible to human error when processing large numbers of seed samples. Recent advances in deep learning and object detection provide opportunities to automate these tasks using conventional RGB images. This study focused on a deep learning-based framework for wheat-seed detection and detection-based counting using a custom red-green-blue (RGB) image dataset. A dataset comprising 832 RGB images containing 3,649 manually annotated wheat-seed instances was developed, with images representing one to ten detached wheat seeds per image. All seed instances were manually annotated using bounding boxes and formulated as a single-class object detection problem. Two lightweight object detection models, YOLOv8n and YOLO11n, were trained and evaluated under identical experimental conditions. The model performance was assessed using precision, recall, mean Average Precision at an Intersection over Union (IoU) threshold of 0.5 (mAP@0.5), mean Average Precision averaged across IoU thresholds from 0.5 to 0.95 (mAP@0.5:0.95), training loss curves, confidence-based performance curves, confusion matrices, and qualitative detection outputs. Both models achieved excellent detection performance on the custom wheat seed dataset. YOLOv8n achieved a precision of 0.9947, recall of 0.9963, mAP@0.5 of 0.9940, and mAP@0.5:0.95 of 0.5284. YOLO11n produced slightly higher performance, achieving a precision of 0.9988, recall of 0.9984, mAP@0.5 of 0.9950, and mAP@0.5:0.95 of 0.5385. Overall, the performance of the two models was similar at mAP@0.5, but at mAP@0.5:0.95 stricter criterion, YOLO11n consistently demonstrated the strongest overall performance. The findings show that lightweight YOLO models combined with RGB imaging provide an effective and easy way for automated localization and counting of wheat seeds. The framework provides a manually annotated RGB wheat-seed dataset and a reproducible benchmark to compare lightweight YOLO models for detection. This study provides a practical foundation for future research on automated seed phenotyping, with future work focusing on external validation using more diverse datasets, multi-class seed-quality assessment, and quantitative evaluation of counting performance.
Christopher Mataczynski · Maike Glaser · Stuart Huntley · Penelope I. Higgs
Environmental bacteria have abundant signaling systems wired into complex gene regulatory networks to adapt to fluctuating conditions. In Myxococcus xanthus, starvation triggers a developmental program (specialized biofilm) that produces spore-filled multicellular fruiting bodies surrounded by a distinct quiescent state termed peripheral rods. Fruiting body structure as well as the proportion of cells following each fate can be tuned by a large repertoire of signaling proteins, including numerous orphan histidine kinases. Here, we focus on the histidine kinase TodK, which was previously demonstrated to influence developmental progression. We find that loss of TodK produces distinct developmental phenotypes that vary with environmental conditions. To quantify these effects, we developed an image-analysis pipeline that measures aggregation and fruiting body patterning during development on nutrient-limited agar. These analyses revealed the todK mutant prematurely aggregates particularly at the peripheries of the colony. Under submerged-culture conditions, initial production of aggregates was observed with wild-type timing, but these aggregates displayed accelerated transition to mature fruiting bodies. Overexpression of active TodK completely blocked fruiting body formation. Molecular analyses demonstrated that TodK overproduction suppressed expression of core developmental regulators including FruA and CsgA. Interestingly, protein accumulation of MrpC, a transcription factor necessary for expression of both FruA and CsgA, was not significantly perturbed. These data suggest TodK silences MrpC's ability to activate transcription of key developmental targets. Together, these findings establish TodK as a modulator of developmental progression and demonstrate how quantitative phenotyping approaches can reveal biologically meaningful functions for orphan histidine kinases whose mutant phenotypes might otherwise appear subtle.
Benoit DE SELON · Jean‐Eudes Hollebecq · Jordan Bernigaud-Samatan · Caroline Chaynes · Solenne Faul-Godec · Mathieu MARGUERIE · Afonso Ponce · Mario Serouart · Samuel Thomas · Tania ROUGIER
CountingSegmentation
The characterization of plants and their environment is changing scale with the development of digital plant phenotyping tools and methods. Phenotyping and envirotyping tools are more accessible, enabling multi-scale observation of plants and their environment—even molecules—through multi-plot observation via satellite. At the same time, large amounts of data are more accessible and easier to exploit through indexing in open information systems. Processing is facilitated by an increasing number of sophisticated tools, particularly with the help of artificial intelligence, thus facilitating analysis by the operator. Two applicated examples are detailled here regarding segmentation for combined mixed crops and insects identification and counting. Environmental and phenotypic data can be combined in predictive or decision support models. These data are valuable assets for accelerating the deployment of agroecology and resilient, sustainable agriculture.
Abstract Rain-induced cracking is among the most damaging preharvest disorders of sweet cherry because even narrow cuticular fractures reduce fresh-market value and create entry sites for water loss and fungal infection. Conventional inspection is subjective and is least reliable for low-contrast microcracks near the pedicel cavity. This study developed a Fourier-guided multi-scale Vision Transformer (FGM-ViT) for joint fruit-level severity grading and pixel-level crack segmentation. The dataset comprised 960 individual fruit from four commercial cultivars and eight orchard blocks, imaged from four rotational views after natural rainfall exposure or controlled water-immersion challenge. Crack severity was assigned as intact, microcrack, moderate, or severe using stereomicroscopic measurements of cumulative crack length, maximum width, affected surface area, and crack location. All views of an individual fruit and all observations from the same orchard block were constrained to the same fold in a nested five-fold grouped evaluation. FGM-ViT combined a multi-scale spatial encoder with luminance-normalised Fourier residual tokens and frequency-guided cross-scale attention. The complete model achieved 93.8 ± 0.9% fruit-level accuracy, 93.7 ± 1.0% macro-F1, and 0.842 ± 0.018 Dice coefficient for crack segmentation. It exceeded the spatial-only transformer by 2.2 percentage points in accuracy and 2.2 points in macro-F1, with the largest gain observed for microcracks. Classification errors were restricted almost entirely to adjacent severity categories. Under pedicel overlap, glare, brightness shifts, and motion blur, FGM-ViT retained a 2.7–3.5-point advantage over the spatial-only model. Frequency-band occlusion indicated that mid-to-high spatial frequencies carried complementary evidence for narrow fractures, whereas spatial features remained essential for crack location and marketability interpretation. The dual-output framework provides an auditable, low-cost RGB approach for sweet-cherry sorting, cultivar screening, and rain-cracking phenotyping.
Background and Aims: Pollen germination and tube growth are critical stages of plant reproduction that are highly sensitive to temperature but remain labor-intensive to quantify. This study aimed to develop a deep learning-based approach for pollen phenotyping and to characterize the temperature dependence of pollen germination and elongation in Douglas-fir (Pseudotsuga menziesii). Methods: A convolutional neural network (CNN) was trained to segment pollen grains and classify germination status from microscopic images. Germination percentage and pollen length were quantified across a range of incubation temperatures from 5 to 40°C. Gamma functions were fitted to temperature response curves to estimate the optimal temperatures for pollen germination and elongation among Douglas-fir populations collected across an elevational gradient. Key Results: The CNN achieved high segmentation accuracy (intersection over union = 0.846), accurately distinguishing pollen grains from the background but showing moderate accuracy in separating germinated from ungerminated pollen during early elongation. Both pollen germination and elongation exhibited bell-shaped temperature response curves with distinct thermal optima. Pollen elongation consistently reached its optimum at higher temperatures than pollen germination. Estimated optimal temperatures fell within a narrow range, approximately 19 to 23°C. No significant relationship was detected between elevation and thermal optima, although some higher-elevation populations exhibited lower optimal temperatures. Comparisons with previous analyses of three western North American conifers showed that each species occupied a distinct reproductive thermal niche corresponding to the spring temperatures of its native habitat. Conclusions: Deep learning provides an efficient approach for high-throughput quantification of pollen germination and elongation from microscopic images. The narrow thermal range for reproductive performance suggests that Douglas-fir pollen is sensitive to temperature variation and that warming climates may alter reproductive success. These findings improve our understanding of the thermal sensitivity of conifer reproduction and provide a scalable framework for assessing impacts of climate warming on forest regeneration.
Classification2D/3D reconstructionWater status / transpiration
Abstract Monitoring pollen is important for crop productivity, plant breeding, and environmental forecasting, yet current approaches remain limited in throughput and physiological insight. Here, we combine single-shot plasma emission imaging with artificial intelligence for pollen morphology reconstruction and apparent hydration state classification. Plasma plumes generated during laser–pollen interaction provide indirect measurements of pollen morphology, extending plasma-based analysis beyond conventional spectroscopic readouts. A conditional generative adversarial neural network reconstructed morphology from such plasma images, while chromaticity analysis revealed systematic color shifts between hydrated and dehydrated grains. A support vector machine using chromaticity histogram and scalar color features distinguished these groups with ∼76% leave-one-out accuracy. The results suggest that plasma emission contains structural and hydration-associated information, supporting rapid pollen morphology inference and physiological phenotyping for agricultural and environmental monitoring.
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.
Accurate point-cloud segmentation of cotton organs is essential for precise phenotypic characterization. However, reliable instance segmentation of cotton bolls in field-derived point clouds remains challenging because foliage occlusion, contact between adjacent bolls and incomplete reconstruction obscure instance boundaries. Here we present RQ-PointNeXt, an end-to-end framework that directly maps input point clouds to boll instance masks within a unified trainable network. Built on PointNeXt, it incorporates relative-elevation geometric channel attention in the shallow encoder to fuse global channel context with elevation and surface-normal cues. Its Query–mask branch integrates semantic guidance, center-seeded queries and a center-aware mask prior to suppress background responses, localize instances and constrain mask extent. Hungarian matching and multitask optimization establish one-to-one query–instance assignments, whereas query-based decoding produces instance masks without external geometric clustering. We evaluated the framework on 226 field-grown cotton plants containing 720 annotated boll instances reconstructed from UAV multi-view imagery using neural radiance fields. On the held-out test set, overall accuracy, mean class accuracy and mean intersection over union reached 0.8942, 0.8980 and 0.8076, respectively. AP25, AP50 and AP75 were 0.7359, 0.5585 and 0.2777, yielding an mAP25/50/75 of 0.5240. The framework provides instance-level outputs for boll counting and spatial analysis in high-throughput field phenotyping.
Reproduction assets foundThe paper's own field cotton point-cloud dataset (226 plants, 720 annotated bolls) is only available upon request. However, the authors directly used the public UGA-BSAIL Cotton Plants with Foliage point-cloud dataset (with their added boll instance annotations) as an evaluation benchmark for RQ-PointNeXt, and it is公开发Dataset · publict to the pointwise overlap between
predicted and ground-truth instances.
To further evaluate the proposed method under conditions of relatively high point-
cloud completeness, experiments were conducted using the public UGA-BSAIL Cot-
ton Plants with Foliage dataset. The point-cloud data are publicly available through
Figshare (https://figshare.com/projects/Cotton_plant_with_foliage/258065, accessed on
8 September 2026), while the associated code and documentation are hosted on GitHub
(https://github.com/UGA-BSAIL/Cotton_plants_with_foliage, accessed on 8 September
2026). The dataset contains relatively complete cotton plant point clouds, surface-normal
attributes, and semantic labels distinOpen asset ↗Figshare · Cotton_plant_with_foliage/258065pdf-raw-page:16 lines:1-52Plant phenotyping relevance matchCrossref · checked 15 Sept 2026
Geospatial artificial intelligence (GeoAI) integrates multimodal remote sensing with deep learning to model complex agricultural systems at scale. Within this framework, accurate and non-destructive prediction of seed composition from in-season standing crops is essential for breeding and precision agriculture. This study developed an end-to-end convolutional neural network (CNN) framework to estimate eight seed traits (protein, oil, sucrose, fiber, starch, ash, complex and simple carbohydrates) from UAV-based multisensor imagery and associated genotype and phenological metadata. A total of 372 soybean samples were collected over two growing seasons (2020–2021) from two fields in Missouri, with UAV flights capturing multispectral (MSI), thermal (THR), and LiDAR (LDR) data at four time points spanning vegetative to reproductive growth stages. CNN models were trained in single- and multi-date configurations, incorporating genotype (GEN) and days after sowing (DAS) as additional features. The highest accuracy was achieved for sucrose (R2 = 0.80), followed by simple carbohydrate (R2 = 0.70) and starch (R2 = 0.55), with notable gains from GEN and DAS. Multi-date models incorporating earlier acquisitions often matched or outperformed later or all-date combinations. Among modalities, MSI provided the most robust estimates, with limited added value from LDR or THR. Unlike feature-based pipelines prone to multicollinearity, this image-to-trait approach enables automated, scalable prediction of soybean seed composition for in-season, field-level assessment.
Lois Onyejere Nwobodo · Chinonso J. Okonkwo · Edith Angela Ugwu · Udeh Chukwuma Callistus · Okorie Kingsley Maduabuchi · Ngene John Ndubisi · Agbo Kenechukwu Martin
Field / plotLeafClassificationStress / disease detectionDisease symptoms / severity
Plant diseases threaten agricultural productivity, and automated image analysis can support early identification of visible disease symptoms. This study introduces HybOptic-CNN, a convolutional neural network (CNN) whose learning rate and batch size are selected using a hybrid Whale Optimization Algorithm--Grey Wolf Optimizer (WOA-GWO). Nine disease classes were selected from the 22-class CCMT field-image dataset, and 152 local farm leaf images were collected in Enugu State, Nigeria. Of the local images, 122 (80.3%) were added to the model-development data for training and validation, whereas 30 (19.7%) formed an independent Nigerian hold-out set excluded from augmentation, class balancing, early stopping, validation, and hyperparameter selection. Across 10 model-development runs, the optimized model achieved 96.8 ± 0.4% mean validation accuracy, 95.2 ± 0.5% macro-precision, 94.9 ± 0.6% macro-recall, and 95.0 ± 0.5% macro-F1, compared with 90.3 ± 0.9% validation accuracy and 86.3 ± 1.2% macro-F1 for the baseline. The optimized model improved mean validation accuracy by 6.5 percentage points and converged 14.6 epochs earlier. On the independent 30-image Nigerian hold-out, HybOptic-CNN achieved 93.3% accuracy and 93.1% macro-F1 across four represented disease classes. A web application integrating the trained classifier was also demonstrated. These results support improved model-development performance through hybrid hyperparameter selection and motivate broader multi-location field evaluation.
Reproduction assets foundThe paper's Data availability statement points to two public sources: a Mendeley dataset (the locally collected Nigerian field images) and the Kaggle CCMT plant disease dataset used as the principal image source. Only the Kaggle URL matches an allowed URL; the Mendeley URL is not in the allowed list, so only the CCMT/KDataset · publicnt and independent field-test data and should pub-
lish the class-wise split manifest, random seeds, WOA-GWO
numerical settings, and evaluation code so that the reported pro-
cedure can be reproduced and extended.
Data availability
The data used in this study are available at https://
data.mendeley.com/datasets/bwh3zbpkpv/1 and https://www.kaggle.com/datasets/rahimanshu/ccmt-plant-disease-dataset.Declaration of competing interest
The authors declare that they have no known competing fi-
nancial interests or personal relationships that could have ap-
peared to influence the work reported in this manuscript.
Funding
The authors received no specific funding from any public,
commercial, or not-fOpen asset ↗Kaggle · rahimanshu/ccmt-plant-disease-datasetpdf-raw-page:12 lines:1-78Plant phenotyping relevance matchCrossref · checked 15 Sept 2026
The precise temporal characterization of early growth dynamics in crops under abiotic stress is critical for stress resistance management in smart agriculture. However, traditional manual measurement methods struggle to achieve high-throughput and accurate quantification of multiple phenotypic parameters over continuous time series. To address this, we developed a full-time-series crop growth monitoring system that continuously collects fine-grained image data of cucumber seed germination and seedling growth under soil culture conditions, constructing an annotated dataset for identifying germination status and quantifying cotyledon area. The YOLOv8n-F_SGX and YOLOv8-seg models were developed and deployed, achieving detection and segmentation accuracies of 0.978 and 0.987, respectively, which enabled automatic monitoring of germination rate and precise extraction of cotyledon area. To further investigate the mitigating effects of nanomaterial priming on salt stress, we conducted germination and seedling growth experiments on cucumber seeds using a series of ZnO-NPs (particle size 30 nm) suspension concentrations at different salt stress levels. The results demonstrated that under salt stress conditions of 0–150 mmol·L −1 , priming with 100 mg·L −1 ZnO-NPs resulted in optimal germination dynamics (highest germination rate and fastest germination speed), along with the largest cotyledon area and highest growth rate in seedlings. In contrast, when the ZnO-NPs concentration increased to 400 mg·L −1 or higher, it significantly inhibited seed germination and seedling growth.
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.
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.
Reliable prediction and interpretation of photosynthetic rate are important for precision environmental management in greenhouse tomato production, but model stability is often limited by variable redundancy, measurement-session effects, and environmental heterogeneity. This study developed an integrated gas-exchange-data-based framework for key-factor selection, photosynthetic-rate reconstruction, cross-session validation, environmental correction, and physiology-informed limitation screening. Core predictors were first identified from high-dimensional gas-exchange variables using K-means clustering and random-forest importance analysis. A CatBoost–ExtraTrees–RBF-SVR stacked ensemble was then constructed to reconstruct the leaf photosynthetic rate from selected gas-exchange variables, and its cross-session generalization was evaluated using nested cross-validation and leave-one-file-out (LOFO) extrapolation. Environmental correction was further applied to improve cross-session comparability, and rule-based limitation screening was used to classify potential environmental constraints associated with reduced photosynthesis. The stacked model achieved an RMSE of 0.806 and an R2 of 0.9918 under nested cross-validation, and an RMSE of 0.814 and an R2 of 0.9917 under LOFO extrapolation. After environmental correction, the cross-session variability of the environmentally corrected photosynthetic indicator was reduced by 96.0%, reflecting improved cross-session comparability. Under deployable sensor inputs, performance decreased markedly (A1: RMSE = 4.578, R2 = 0.735; A2: RMSE = 4.610, R2 = 0.731), compared with the gas-exchange baseline (RMSE = 0.833, R2 = 0.991). Thus, routine greenhouse sensor models should be regarded as approximate rather than high-accuracy substitutes for the gas-exchange-based model.
Reproduction assets foundThe paper analyzes a publicly available greenhouse tomato gas-exchange dataset (Manjarrez-Sanchez & Martinez-Carrillo, Data Brief 2020), which is the exact phenotype/sensor input data for this study's photosynthesis reconstruction analysis. The authors' own processed data and analysis scripts are only available upon请求,Dataset · publicThe data used in this study were obtained from the publicly available greenhouse
tomato gas-exchange dataset reported by Manjarrez-Sanchez and Martinez-Carrillo [29].Open asset ↗pdf-raw-page:4 lines:1-38Plant phenotyping relevance matchEurope PMC · checked 15 Sept 2026
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 Image-based plant-disease classifiers often achieve high accuracy under controlled conditions, but their reliability can deteriorate in real-world field environments. This article investigates crossdataset strawberry leaf-health classification using controlled PlantVillage images and field images collected through the SmartBerry project in Nigeria. Four deep-learning architectures were evaluated under within-domain, controlled-to-field and pooled-training conditions using healthy-leaf and darkspot/ scorch symptoms. Although all models achieved near-ceiling performance on PlantVillage, direct transfer to SmartBerry reduced macro F1-score to between 0.418 and 0.824, demonstrating a substantial gap between controlled-dataset performance and field reliability. To reduce the amount of locally labelled field data required to address this gap, this paper proposes SmartBerry Domain Adaptation (SmartBerry-DA). This label-efficient adaptation approach combines a model trained on controlled images with a small set of labelled SmartBerry images and additional unlabelled field images. Using MobileViTv2-0.5, SmartBerry-DA achieved a macro F1-score of 0.940 ± 0.015 with only five labelled SmartBerry images per class and 0.988 ± 0.012 with ten, compared with 0.994±0.008 when the full labelled target dataset was used. This shows that strong field performance can be recovered with substantially reduced local annotation. For lightweight offline deployment, 16- bit floating-point representation reduced model size from 4.37 MB to 2.30 MB without changing predictive performance under the evaluated conditions. The findings demonstrate the importance of locally acquired field data and label-efficient adaptation for transferring agricultural AI from benchmark datasets to practical farming environments, while providing a pathway towards lightweight strawberry disease recognition in settings where extensive annotation and continuous connectivity may not be feasible.
Reproduction assets foundThe paper's own SmartBerry field strawberry image dataset is explicitly stated to be publicly available on Zenodo with a Kaggle mirror. PlantVillage is a cited third-party benchmark, and no author analysis code or trained model checkpoints are stated as available.Dataset · publicip with the SmartBerry partners.
Ethics statement
NA
Data availability
The PlantVillage dataset is publicly available from its
cited source. The SmartBerry dataset is publicly avail-
able through Zenodo, with a Kaggle mirror provided for
convenient access.SmartBerry dataset: Zenodo: https://
zenodo.org/records/22056372; Kaggle: https://www.kaggle.com/datasets/isiborihianle/smartberry-strawberry-field-image-dataset
Funding
This research was supported by the SmartBerry Project
(Project Number: 10071867) and the Department of Com-
puter Science, Nottingham Trent University.
Acknowledgements
The authors thank the Ashley Strawberry Farm staff who
supported image collection, annotation and field eOpen asset ↗Kaggle · isiborihianle/smartberry-strawberry-field-image-datasetpdf-raw-page:17 lines:1-89Plant phenotyping relevance matchCrossref · checked 15 Sept 2026
Published9 Sept 2026Journal of Applied Science and Technology Trends
This study proposes an integrated hybrid pipeline approach which is the combination of an optimized Inception-v4 Convolutional Neural Network (CNN), YOLOv8 (You Only Look Once version 8) and extreme gradient boosting (XG Boost) models to detect and classify Tomato plant diseases effectively and accurately. The proposed methodology is based on the fine-grained localization ability of the YOLOv8 model to accurately localize the affected area of the leaf, multi-scale deep feature extraction ability of the optimized Inceptionv4 CNN, and XG Boost to reduce dimensions and optimize features. The hybrid model combines the Inception-v4 CNN, YOLOv8, and XG Boost models with an average CPU processing time of about 200 milliseconds per image, the model performs better in terms of computational efficiency, robustness, and faster prediction. Experimental tests on Plant Village show that the model outperforms the state of the art in 10 disease categories. Specifically, the integrated hybrid system achieved high precision, recall, F1-score, mean Average Precision (m AP), and its training accuracy was more than 0.98 (up to 0.9969), and the testing accuracy was more than 0.96 (up to 0.9695). The hybrid framework shows good reliability even for diseases that are visually similar, such as bacterial spot, early blight, late blight, mosaic virus and yellow leaf curl virus. The optimization of batch size to 32, along with the tuning of the learning rate further improved the stability of training, convergence speed and the generalization of the overall model. The proposed system holds promise for real-time, scalable, and sustainable precision agriculture, aiming for early disease detection and yield protection. Future work includes incorporating Internet of Things (IoT) edge devices, in field environmental monitoring, multimodal data integration and applying Explainable Artificial Intelligence (EAI) techniques to enhance model interpretability for end-users.
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.
Abstract Tomato leaf diseases (TLDs) such as Early Blight (EB), Late Blight (LB), and Leaf Mold (LM) have a negative impact on crop yield and quality, result-ing in economic losses in agriculture. Traditional diagnosis methods depend on the visual recognition of a disease, which are time-consuming, subjective and may be difficult to reach in rural settings. Keeping these drawbacks in mind, the present work introduces a Transfer Learning model for tomato leaf disease classification based on EfficientNetB3 network. A pretrained Ef-ficientNetB3 network, trained on ImageNet, is used to obtain discriminative features like lesion boundaries, discoloration, fungal textures, and infection spots. Images are cropped to 224 × 224 pixels and split into training, validation and test set. The proposed architecture employs Global Average Pooling, a dense layer with ReLU activation, drop out regularization and Softmax classifier for classification of tomato leaf images into four classes: Early Blight, Late Blight, Leaf Mold and Healthy. Experimental results show the high classification accuracy and low training, validation and test-ing losses. Inter-class misclassifications of the confusion matrix show very few, which supports good generalization. Moreover, Grad-CAM visualiza-tions can generate interpretable heat maps that point to the regions of the image affected by the disease, and multi-class ROC analysis gives high AUC values, which means that it has excellent class separability. The proposed framework provides a precise, reliable, and interpretable approach for auto-mated tomato leaf disease diagnosis, contributing to precision agriculture by assisting in early detection and prompt management of tomato diseases.
Reproduction assets foundThe paper's tomato leaf disease classification uses a publicly available Kaggle dataset (PlantVillage-derived tomato leaf images, 4000 images across 4 classes), explicitly declared in the Data Availability statement with a public URL. No author code, trained models, or other paper-specific assets are disclosed.Dataset · publicitted in accordance with the Journal policies.
Permission to use third-party material
Images or figures are never published previously and did not take from any
internet resources.
Data Availability
The dataset used in this study is publicly available from the PlantVillage
tomato leaf disease dataset on Kaggle’s following link:
https://www.kaggle.com/datasets/kaustubhb999/tomatoleaf
Acknowledgements
The authors would like to thank SR University, Warangal, Telangana,
INDIA and Sharda University, Greater Noida, Uttar Pradesh, INDIA for
providing research facilities and computational resources to carry out this
work.Open asset ↗Kaggle · kaustubhb999/tomatoleafpdf-raw-page:20 lines:1-18Plant phenotyping relevance matchEurope PMC · Crossref · checked 15 Sept 2026
Published8 Sept 2026Springer Science and Business Media LLC
Abstract Jute is one of the most important cash crops in Bangladesh and plays a vital role in the country's economy. However, the yield and quality of jute are often affected by various plant diseases, which are difficult to detect manually at an early stage. Early identification is essential to prevent large-scale damage and ensure sustainable production. This study proposes an automated jute disease detection system using image processing and deep learning techniques. Several models, including a Custom Convolutional Neural Network (CNN), VGG16, DenseNet121, and a Hybrid EfficientNetB7 architecture, were implemented and compared. The proposed Hybrid EfficientNetB7 model achieved the highest classification accuracy of 99.7%, outperforming the other models. The system effectively distinguishes between healthy and diseased jute leaves and stems, providing a reliable tool for early disease detection. This research demonstrates that deep learning-based ensemble and transfer learning methods can significantly improve the accuracy and reliability of plant disease detection systems. The findings can contribute to the development of intelligent agricultural tools, empowering farmers with faster and more accurate disease diagnosis for better crop management and yield improvement.
Abstract DenseNet121 is effective for rice disease recognition, but its dense feature reuse keeps many connections active during inference. We evaluated whether learnable layer gates could improve paddy disease classification and expose removable DenseNet connections. Using the Kaggle Paddy Disease Classification dataset, we trained an ImageNet-pretrained DenseNet121 baseline and an Adaptive Gated DenseNet121 on 8,325 training images and validated on 2,082 images across ten classes. The baseline achieved 96.78% accuracy and 96.64% macro F1-score. The adaptive model achieved 97.21% accuracy and 96.89% macro F1-score, with 97.16% macro precision and 96.65% macro recall. Pruning at threshold 0.825 reduced parameters from 6.96 million to 6.84 million while retaining 96.39% F1-score, but slightly higher thresholds caused severe collapse. Adaptive gating therefore improved classification and enabled limited pruning, but did not yet provide robust mobile compression.
Reproduction assets foundThe paper uses the public Kaggle Paddy Disease Classification image dataset for all phenotyping measurements and states that the authors' PyTorch implementation, Kaggle notebooks, and report-generation scripts are publicly available in a GitHub repository. Both are paper-specific, public, and actionable.Dataset · publicarticipants, human tissue, live vertebrates or higher invertebrates.
Use of artificial intelligence assistance
A large language model was used to assist with code organization and formatting.
Data availability
The dataset presented and used in this study are the publicly available Paddy Disease Classification dataset on Kaggle: https://www.kaggle.com/competitions/paddy-disease-classification/data.
Code availability
The PyTorch implementation, Kaggle notebooks and report-generation scripts are available in the project workspace on Github [20].
Declarations
Acknowledgements
The authors acknowledge the Kaggle Paddy Disease Classification competition and the Paddy Doctor dataset contributors forOpen asset ↗lines:177-215Code · publicThe PyTorch implementation, Kaggle notebooks and report-generation scripts are available in the project workspace on Github [20].Open asset ↗lines:177-215Plant phenotyping relevance matchOpenAlex · arXiv · checked 15 Sept 2026
Daniel Rosendo · Renan Souza · Kelsey Carter · John Lagergren · Frédéric Suter · Shelaine Curd · David Weston · Rafael Ferreira da Silva
Laboratory / benchtop
Autonomous, cross-facility science requires capabilities that no individual project should have to build for itself: managed execution for long-lived services, versioned distribution of models to remote compute systems, governed access to large language models, a shared substrate for experimental data, and end-to-end provenance. The U.S. Department of Energy Genesis Mission platform, delivered through the American Science Cloud, provides these as reusable services. This paper reports how the Genesis platform enables cross-facility experiments and accelerates scientific discovery. We explore the plant phenotyping workflow of the Orchestrated Platform for Autonomous Laboratories as the exemplar: it couples Oak Ridge National Laboratory's Advanced Plant Phenotyping Laboratory with the Frontier supercomputer. In a 40-day nickel-treatment campaign, the resulting workflow replaced roughly twelve hours of manual analysis with interactive queries returning in seconds to minutes.
Abstract Background and Aims Drought reduces wheat yields, yet field-scale quantification of root water uptake (RWU) remains challenging because below-ground processes are difficult to monitor. This study developed a non-invasive hydrogeophysical framework integrating Electrical Resistivity Tomography (ERT), TDR-based soil monitoring, and depth-aware Random Forest calibration to quantify depth-resolved RWU and evaluate genotype-specific water-use strategies under terminal drought. Methods Time-lapse ERT (44 surveys, ≥ 3 week⁻ 1 ) was combined with TDR sensor measurements of soil water content (n = 278 paired ρ–θ observations) to convert resistivity measurements into depth-resolved RWU estimates across 0.1–1.0 m depth. Five petrophysical models were evaluated using date-grouped fivefold cross-validation, with the depth-aware Random Forest performing best. Three wheat genotypes with contrasting root architectures were monitored under terminal drought (142 mm available water). ERT-derived RWU were analysed alongside stomatal conductance, chlorophyll fluorescence, and grain yield. Results ERT resolved RWU strategies among genotypes. WM-203 exhibited aggressive, coordinated multi-layer water extraction across the soil profile (r = 0.80–0.98), whereas WM-140 showed a delayed uptake strategy characterized by early deep-layer dominance followed by mid- and deep-profile engagement, and IPLR-760 displayed inconsistent uptake with mid-profile hydraulic decoupling. Genotypic RWU rankings were consistent with stomatal conductance and grain yield, spanning from 7.0 t ha⁻ 1 in WM-203 to 1.5 t ha⁻ 1 in IPLR-760 despite comparable total water extraction. Conclusion ERT-based quantification of RWU provides a robust, non-invasive approach for resolving genotype-specific water-use strategies under field conditions. The framework enables characterization of water-use coordination patterns and offers a tool for phenotyping drought-resilient wheat genotypes.
Reproduction assets foundThe paper's Data availability statement explicitly states that the code and supporting data for this ERT-based root water uptake study are publicly available on the authors' GitHub repository, which is listed in allowed_urls. This qualifies as a paper-specific public code/data asset for the phenotyping analysis.Code · publicsity of Jerusalem. This research was supported by the Chief
Scientist of the Israeli Ministry of Agriculture and Food Secu-
rity (grant no. 12–01-0056) and the Israeli Council for Higher
Education (Project: Future Crops for Carbon Farming).
Data availability The code and supporting data for this study
are publicly available at:
https://github.com/emmaiyke/ERT_RWU_Wheat_Project
Additional datasets are available from the corresponding
author upon reasonable request.
Declarations
Competing interests The authors declare that they have no
known competing financial interests or personal relationships
that could have appeared to influence the work reported in this
paper.
Open Access This article isOpen asset ↗ERT_RWU_Wheat_Project · emmaiyke/ERT_RWU_Wheat_Projectpdf-raw-page:22 lines:1-95Plant phenotyping relevance matchEurope PMC · checked 15 Sept 2026
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.
Plant diseases, being a subject of interdisciplinary research, significantly reduce crop yield, quality, and economic returns, while the misidentification of pathogens often leads to ineffective treatments and may harm beneficial organisms and ecosystems. This work develops an approach for robust visual classification of plant diseases under limited and heterogeneous data based on multi-scale fractal texture descriptors integrated into a convolutional neural network. The proposed method employs wavelet transform modulus maxima to extract two complementary fractal characteristics, local fractal dimension and singularity spectrum width, from leaf images at several spatial scales. These descriptors form multi-channel fractal maps fed into a fractal attention module (FAM) inserted after the third stage of a ResNet-50 architecture. The FAM learns to emphasize spatial regions where fractal properties are most discriminative, while a parallel branch encodes global fractal statistics into an auxiliary vector combined with backbone features at the final classification layer. Experiments are conducted on a large heterogeneous collection of 11 public plant disease datasets under 5-shot, 50-shot, and full-scale training regimes. The fractal-augmented model raises classification accuracy from 57.06% to 67.73% on 5 shots and from 80.81% to 86.11% on 50 shots, red outperforming the plain ResNet-50 in these settings, converges within 1–2 epochs versus 25–40, and shows markedly better resilience to color distortions, random occlusions, and grayscale conversion in most cases. The generated attention maps provide spatially explicit explanations of the model’s decisions, increasing transparency for practical use. The proposed approach demonstrates that fractal analysis, embedded as a modulating signal inside a deep network, can serve as an efficient and interpretable inductive bias, which is particularly valuable under data scarcity and noisy agricultural imagery.
Field / plotFlowerClassificationObject detectionGrowth / development / phenology
Abstract Dioecious crops face significant pollination challenges due to the asynchrony in flowering between male and female plants. This asynchrony varies spatially across orchards, requiring targeted interventions in zones where synchrony is lacking. Assisted pollination addresses this deficiency, albeit at a substantial operational cost that could be optimised through spatial phenological mapping. Manual assessment proves economically infeasible at commercial scales, while existing computer vision systems are unable to classify phenological stages and integrate geospatial information. This study presents the Mobile Phenological Mapping (MPM) Framework, which integrates GNSS-synchronised smartphone video with automated phenological detection to generate plant-level phenological distribution maps validated in commercial kiwifruit (Actinidia chinensis) orchards. This modular framework comprises (i) training and validation data acquisition, (ii) model optimisation, (iii) operational pipeline, and (iv) performance evaluation. MPM employs hierarchical deep learning across three stages: structure detection, gender classification, and phenological stage classification. Video frames are georeferenced through timestamp matching with GNSS metadata, enabling spatial phenological mapping. Operational validation across four commercial orchard zones demonstrated mean absolute percentage errors of 17.2% for structure detection and 20.1% for gender classification. The framework reduces monitoring time from 113 to 1.6 hours per hectare, decreasing labour costs from €2 060 (113 hours × 18.20 € per hour) to €29 (1.6 hours × 18.20 € per hour) per hectare based on the Portuguese hourly labour cost for minimum wage workers. When integrated with routine orchard operations, video acquisition incurs negligible additional cost. MPM provides growers with precision phenological maps for targeted pollination interventions. While validated in a kiwifruit orchard, the modular architecture can be adapted to other crops by replacing the training data.
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.
Abstract The timely detection of plant health status is essential to obtaining several advantages, such as increasing crop yield, reducing the toxic pollutants used in crops, enhancing healthy crops, and enhancing economic returns. A computer-aided plant status identification system enables plant health identification using plant leaves because plant leaves are the most prominent in the plant. Deep learning (DL) and convolutional neural networks (CNN) are distinguished in the area of plant health identification. Still, CNN cannot handle images of variable size and fails to extract features efficiently if the image has complex backgrounds. To avoid these problems, this research proposes a hybrid model for plant (groundnut) illness status identification using the CNN and vision transformer (HCVT). CNN is efficient in identifying local features, and ViTs are efficient in identifying global features from leaf images; leveraging these advantages enhances the model's efficacy. The HCVT model used the two datasets, groundnut and rice datasets, for experimentation. The HCVT model obtained an accuracy of 96.40% on the groundnut leaf image dataset, with 99.86% accuracy on the rice dataset. The ablation analysis performed the necessity of each component in the HCVT model. The LIME (local interpretable model-agnostic explanations) technique is utilized to comprehend the HCVT method functionality and its results showed that the HCVT model accomplished better than the present cutting-edge models in identifying plant illness and exhibiting its generalization potential. The HCVT model is affordable and widely accessible for analysing plant leaf images. Utilizing the HCVT model offers a robust and easily accessible approach for diagnosing plant diseases by analysing leaf images.
Rice leaf diseases significantly reduce agricultural yield and pose a major challenge for sustainable food production, particularly owing to the limitations associated with manual and visual inspection methods that are subjective and often ineffective in early-stage detection. This investigation proposes a new Hybrid ResConvolutional Neural Network (HyResCN-Net) for effective classification and severity identification of rice plant leaf diseases. The proposed framework integrates an Internet of Things (IoT) -based data acquisition and routing simulation using CrowWhale Energy Trust Routing (CrowWhale-ETR) for efficient data handling. Initially, preprocessing is done by an averaging filter to reduce noise. Then, plant leaves are segmented using the Eff-UNet++ method. Augmentation techniques like rotation, scaling, and color change are applied to expand the dataset. Features, like entropy with Gradient Directional Pattern (GDP), Complete Local Binary Pattern (CLBP), and histogram features, are extracted to enhance feature representation. These features are then used within the proposed HyResCN-Net model, which integrates Parallel Convolutional Neural Network (PCNN) and ResNeXt to improve discriminative learning for disease classification and severity estimation. Experimental evaluation is conducted on the Rice Leaf Bacterial and Fungal Disease Dataset. Considering a k-value of 8, the HyResCN-Net gains an accuracy of 94.258%, a True Positive Rate (TPR) of 96.479%, a True Negative Rate (TNR) of 92.898%, a precision of 91.312% and an F1-score of 93.824% compared to existing methods. The HyResCN-Net efficiently enhances rice leaf disease identification and severity analysis, supporting its applicability in precision agriculture applications.
Abstract Deep learning models have achieved near-perfect accuracy in plant disease classification within controlled laboratory settings; however, their deployment in real-world agricultural environments is severely hindered by the “deployment gap”—a critical vulnerability to environmental corruptions such as sensor noise, motion blur, and occlusion. To bridge this gap, we propose the Recurrent Active Vision Agent (RAVA), formulating the disease detection task as a Partially Observable Markov Decision Process (POMDP). Unlike passive Convolutional Neural Networks (CNNs) that process images globally, RAVA mimics the active inspection behavior of human agronomists. Our architecture integrates a lightweight ResNet-18 backbone with a Recurrent Neural Network (RNN) and a Spatial Transformer Network (STN). Driven by Proximal Policy Optimization (PPO), the agent learns a sequential policy to intelligently navigate and zoom in on informative “glimpses,” effectively bypassing background clutter. To stabilize the reinforcement learning process and enforce noise-invariant feature representations, we introduce a hybrid objective incorporating Supervised Contrastive Learning (SupCon). Comprehensive experiments on a combined PlantVillage and PlantDoc dataset demonstrate RAVA’s overwhelming superiority under extreme conditions. In a “Severe Degradation” stress test, standard ResNet-50 accuracy collapses to 26.3%, whereas our active agent maintains a robust 77.0%. Under extreme noise and occlusion, RAVA preserves 54.6% accuracy compared to the baseline’s 17.2%. Notably, this resilience is achieved with merely ∼12M parameters—significantly fewer than large-scale Vision Transformers—proving that active visual attention, coupled with contrastive learning, offers a computationally efficient and highly robust pathway for field-ready precision agriculture.
Quantitative pollen viability analysis is a critical but labor-intensive step in plant reproductive biology. Existing deep-learning Segment Anything Models (SAM) fail to reliably segment viable pollen in Alexander-stained anthers. To address this, we fine-tuned an existing Cellpose-SAM model for pollen segmentation. We integrated it into PAT (Pollen Analysis Tool), a cross-platform desktop application. PAT features instance segmentation with interactive quality control, an in-app model retraining module, and publication-ready statistical outputs. We deployed PAT in an EMS suppressor screen of semi-sterile Arabidopsis smg7-6 mutants, enabling efficient candidate prioritization for whole-genome sequencing and mapping of the candidate mutation. This screen led to the identification of a point mutation in CAP-D2 (capd2-2), a Condensin I subunit, that rescues the smg7-6 meiotic phenotype. Notably, mutation in a Condensin II subunits (CAP-D3 and CAP-H2) does not confer rescue. Further characterization suggests the capd2-2 allele is hypomorphic, showing no defects in vegetative growth, chromocenter compaction, or transposable element silencing. Collectively, we demonstrate that accessible AI tools have the potential to bridge gaps in plant phenotyping and accelerate the pace of biological discovery.
Reproduction assets foundThe paper's Data availability statement explicitly deposits the authors' PAT pollen-phenotyping tool (the software implementing the paper's computational analysis, including the fine-tuned CPSAM segmentation model support) as open source on GitHub. Note: the full repository URL in the text (https://github.com/Riha-429[Code · public17
Data availability
428
Pollen Analysis tool (PAT) is available as open source tool at Github repository (https://github.com/Riha-429
Lab/Pollen-Analysis-Tool).
430
Figure legends
431
Fig. 1. Cellpose performance on Alexander-stained anther cross-sections across varying pollen
432
densities.
433
Representative cross-sections of Alexander-stained anthers showing a range of pollen densities, from
434
low (top rows, light staining) to high (bottom rows, dense reOpen asset ↗Pollen-Analysis-Toolpdf-raw-page:17 lines:1-64Code / dataset availability confirmedOpenAlex · Crossref · checked 15 Sept 2026
Jie Zhou · Victor Hugo Moura de Souza · Lei Ju · Sebastian Eves‐van den Akker · Ji Zhou · Olaf Prosper Kranse
ArabidopsisRootMorphology / geometry measurementSegmentationRoot system architecture
Introduction Plant parasitism by sedentary plant-parasitic nematodes is a dynamic and continuously evolving process, accompanied by profound remodelling of host root system architecture across distinct infection stages. However, the physiology and anisotropic growth of Arabidopsis thaliana roots under Heterodera schachtii infection, together with complex lateral root proliferation and increasingly dense, overlapping morphology, pose substantial challenges for accurate image segmentation. Methods Here, we introduce Root-TransUNet, an optimised segmentation architecture that extends TransUNet by incorporating a composite loss function to improve boundary precision and structural continuity, as well as dual-stage strip-pooling (SP) modules to enhance elongated and directional root features. These adaptations address the unique morphological complexity of the infected root system. Additionally, we integrated Root-TransUNet into a high-throughput phenotyping pipeline and applied it to an existing dataset of ~120,000 images of 362 A. thaliana MAGIC recombinant inbred lines collected over several months of infection. By extracting root system architecture traits, including root surface area and estimated root volume across infection stages, we enabled stage-specific association analyses between host root growth and nematode performance across these genotypes. Results Root-TransUNet achieved strong segmentation performance, demonstrating improved structural continuity and boundary precision compared with widely used CNN- and Transformer-based baselines, including UNet++. Stage-specific analyses revealed that the relationship between host root traits and nematode performance changed as infection progressed. During establishment, nematode number was largely independent of initial root size and varied strongly among genotypes, whereas during the reproductive phase (10-30 dpi), greater root expansion coincided with reduced estimated nematode volume accumulation. Notably, nematode burden was largely independent of host root size before infection, indicating that root quantity was generally not a limiting factor for infection in this experiment. Discussion These results demonstrate that Root-TransUNet can robustly segment infected root systems across a wide range of nematode infection densities, providing a scalable image-analysis framework for studying plant-parasitic nematode parasitism in combination with host root phenotyping.
Reproduction assets foundThe paper analyzes a public BioImages dataset (S-BIAD2402) of ~400,000 RGB root/nematode infection images and provides authors' analysis code on GitHub; both are paper-specific, public, and actionable.Code · publicng molecular signatures, deepening our understanding of host-parasite resource allocation strategies, and establishing a foundation for the discovery of novel resistance mechanisms.
Code and data availability
Python-based source code for automating root analysis using the datasets above is accessible via our GitHub repository ( https://github.com/JieZhou1025/Root-nematode-interaction ).
Statements
Data availability statement
Publicly available datasets were analyzed in this study. This data can be found here: https://www.ebi.ac.uk/biostudies/bioimages/studies/S-BIAD2402 .
Ethics statement
The manuscript presents research on animals that do not require ethical approval for their study.
AuthorOpen asset ↗JieZhou1025/Root-nematode-interactionlines:412-424Plant phenotyping relevance matchCrossref · checked 15 Sept 2026
Overlapping cells in plant suspension-culture microscopy pose a particular challenge, for instance, segmentation because a single pixel may belong to more than one cell. Most standard instance-segmentation methods are not designed for this setting and tend to treat overlapping objects as mutually exclusive regions. We instead represent each cell as an independent full-cell instance and introduce MC-SlotNet, an architecture that separates competitive object-slot feature assignment from mask decoding. This allows multiple predicted masks to occupy the same image region. We further introduce a mask-level multiplicity-consistency loss that encourages the predicted number of masks covering a pixel to agree with the underlying cell occupancy. We evaluate MC-SlotNet on a newly annotated dataset of 53 Siraitia grosvenorii suspension-culture micrographs containing 4131 full-cell instances acquired at 4×–40× magnification. Using grouped five-fold cross-validation and an overlap-preserving evaluation protocol, we compare the method with Mask R-CNN, SOLOv2, and Mask2Former. MC-SlotNet achieves the best performance on AP50 (0.800), mAP50:95 (0.565), F150 (0.842), all-ground-truth Dice (0.761), AJI+ (0.754), overlap-region Dice (0.705), and overlap-instance recall (0.828). Its AP75 (0.649) is comparable to Mask2Former’s (0.651). MC-SlotNet also has the lowest inference time among the evaluated methods, at 1.113 s/image. These results indicate that decoding full-cell masks independently, rather than enforcing an exclusive partition of image pixels, is well-suited to instance segmentation in plant suspension-culture microscopy images with substantial cell overlap.
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.
Near-infrared imagery is essential for vegetation monitoring, precision agriculture, and environmental remote sensing, but multispectral UAV systems remain significantly more expensive and less accessible than conventional RGB imaging platforms. This study presents a lightweight artificial intelligence framework for reconstructing the NIR spectral band exclusively from the red spectral band acquired by a UAV. The proposed methodology formulates the reconstruction task as a pixel-wise nonlinear regression problem and employs a compact multilayer perceptron (MLP) containing only 609 trainable parameters, without exploiting spatial neighborhood information. The framework was developed and evaluated using 280 synchronized multispectral UAV image sets acquired with a DJI Phantom 4 Multispectral platform over a heterogeneous agricultural landscape in the Republic of Moldova. Of these, 252 image sets were used for model development, and 28 were reserved as a held-out within-mission test subset. Quantitative evaluation on a held-out test dataset from the same acquisition mission yielded a mean squared error of 0.010329, a root mean squared error of 0.101632, a mean absolute error of 0.079883, a coefficient of determination of 0.253383, and a Pearson correlation coefficient of 0.683637 between measured and reconstructed normalized NIR digital intensities. The results indicate that the model captures part of the red–NIR relationship under the evaluated acquisition conditions; however, the moderate coefficient of determination suggests that the reconstructed values are an approximation rather than a replacement for measured NIR observations. An illustrative NDVI-based assessment showed that broad spatial vegetation patterns remained identifiable. Rather than introducing a new neural network architecture, this work establishes a compact empirical baseline to investigate the practical performance and limitations of pixel-wise NIR reconstruction from a single red-band value with minimal model complexity.
Naincy Sagar · Nassim Belmokhtar · Nathalie Boizot · Ana Alves · David Chassagnaud · R Fichot · José Rodrigues · Luc Pâques
Raman / spectroscopyLeafTissuePhysiological trait estimationLeaf traitsWater status / transpiration
Phenotyping extensive populations remains a major constraint in tree breeding programmes, particularly due to the time-consuming and labour-intensive nature of conventional methods. Near infrared (NIR) spectroscopy, which is a high-throughput phenotyping method, offers an alternative solution, providing a rapid and cost-effective approach for assessing growth- and function-based traits on large numbers of trees. This study aimed to evaluate the potential of NIR spectroscopy-based models for predicting such traits in Larch. Specifically, delta carbon-13 ( δ 13 C), carbon (C), nitrogen (N), specific leaf area (SLA), leaf dry matter content (LDM), and phenolics on needles; the branch hydraulic trait (P 50 ), and lignin and hydroxyphenyl/guaiacyl (H/G) ratio on wood cores from an experimental study on Larix species were predicted using multivariate modelling, specifically, partial least squares regression. Reliable models were obtained for N content (R 2 training = 0.95, r 2 testing = 0.94), lignin (R 2 training = 0.95, r 2 testing = 0.94), and H/G ratio (R 2 training = 0.88, r 2 testing = 0.89), while moderate predictive performance was observed for C content (R 2 training = 0.79, r 2 testing = 0.79) and δ 13 C (R 2 training = 0.76, r 2 testing = 0.69). This methodological approach and its results encourage the transition from traditional laboratory methods to efficient, large-scale-based trait evaluation techniques in forestry.
Shail Bala · S. I. Harlapur · Aditya Kamalakar Kanade · M. P. Potdar · Gurupada B. Balol · V. B. Kuligod · Lingareddy Usha Rani
Field / plotLeafClassificationDisease symptoms / severity
Early and accurate detection of plant diseases is critical in precision agriculture to improve crop management and yield. Mungbean ( Vigna radiata L.) is highly susceptible to several foliar diseases, including yellow mosaic, powdery mildew, leaf crinkle, and cercospora leaf spot, which cause substantial productivity losses. Despite expanding applications of deep learning in plant disease diagnosis, systematic multi-architecture evaluation for mungbean disease classification under natural field conditions remains limited. This study addresses this gap by evaluating five state-of-the-art deep convolutional neural network (DCNN) architectures on a large-scale, field-acquired mungbean dataset that captures real-world variability across environmental conditions and disease severity levels, distinguishing it from controlled laboratory studies. A total of 5,617 original images across five classes were used. Data augmentation was applied exclusively to the training subset after stratified splitting to prevent data leakage. The dataset was partitioned into training (70%), validation (15%), and testing (15%) subsets. VGG16, VGG19, ResNet50V2, DenseNet121, and InceptionV3 were evaluated using identical transfer learning and fine-tuning protocols. Model performance was assessed using AUC-ROC, Cohen's kappa coefficient, McNemar's test for pairwise statistical comparisons, five-fold cross-validation, and Grad-CAM-based interpretability. On the independent test set, InceptionV3 achieved the highest accuracy (98.47%) and macro-F1 (98.49%), followed by VGG16 (98.36%) and VGG19 (97.89%). AUC-ROC values exceeded 0.997 for all models, confirming excellent class discrimination. Grad-CAM visualizations further confirmed that model predictions were based on biologically relevant disease symptoms. The findings demonstrate the effectiveness of deep learning for robust disease recognition under realistic field conditions and highlight the potential of AI-based diagnostic tools for crop health monitoring, precision agriculture, and decision-support systems in mungbean production.
Automated fruit counting and yield estimation systems are necessary for efficient orchard management. This study presents a computer vision system based on video multi-object-tracking for fruit load estimation in apple orchards and provides a comprehensive analysis of the system performance under diverse scanning conditions. The system integrates fruit detection, tracking, localization within orchard, and fruit load map generation. Experiments were carried out in an experimental apple orchard containing 420 apple trees. Data was collected with two different RGB-D sensors (Azure Kinect DK and ZED 2) at three different scanning distances (125 cm, 175 cm, and 225 cm) on two different dates prior to the harvest. Comparing the two evaluated sensors, Azure Kinect provided more consistent performance across different dates. Results also show that the longer scanning distance improves accuracy due to seeing the full tree view gives better fruit counts than close partial views. Between the two dates, best results were achieved near harvest due to fruit color at this stage, achieving a Mean Absolute Percentage Error (MAPE) of 6.91 % and a determination coefficient (R 2 ) of 0.733 (using ZED2 sensor at 225 cm distance). Finally, a test comparing scanning from one or both sides of the tree row showed that bilateral scanning improved fruit load estimation at the stretch level by incorporating information from both sides of the canopy. The results of this work demonstrate the effectiveness of the video fruit tracking systems as a useful tool for automating fruit load estimation.
Leaf hyperspectral reflectance can provide a scalable way to estimate photosynthetic capacity (Vcmax), but models trained in one species or measurement context often lose accuracy in another. This transfer problem limits the use of spectral approaches in multi-species crop phenotyping and carbon-cycle applications. Here, we tested physiology-informed inputs for leaf-level Vcmax25 retrieval using paired gas-exchange and reflectance data from wheat (C₃; n = 198) and maize (C₄; n = 81) grown under contrasting nitrogen supply. The four input configurations were raw spectra (Mod1), spectra scaled by a PPFD–absorptance proxy (Mod2), scaled spectra augmented with radiative-transfer-derived traits (Mod3), and scaled spectra augmented with a spectral coordination proxy (Mod4). Within datasets, the best models reached R² = 0.82 in wheat, 0.41 in maize, and 0.76 in the combined dataset. In a matched comparison with a common random-forest learner, the spectral coordination proxy Mod4 improved accuracy only slightly over Mod2 in wheat (RMSE −0.51%; p = 0.0058) and maize (RMSE −1.26%; p = 0.0011) but not in the combined dataset (RMSE −0.15%; p = 0.074), and the trait-based Mod3 showed no consistent benefit. When wheat models were tested on measurement dates not used in training, accuracy remained moderate (R² = 0.563; RMSE = 15.07 µmol m⁻² s⁻¹). Despite this within-dataset performance, models applied to the other species without calibration failed in both directions (negative R²), and adding source-species data did not improve prediction even when a few samples of the new species were used for calibration. These results show that physiology-informed input design provides at most small within-dataset gains, and that reliable prediction across C₃ and C₄ crops requires calibration data from the target crop.
Fei Liu · Yingjie Fan · Mingtao Zhou · Huabing Liu · Shuya Chen · Bin Wen
LeafSegmentation
Abstract Accurate leaf instance segmentation is fundamental to quantifying morphological and structural traits in image-based plant phenotyping. However, substantial variations in leaf scale, shape, and orientation, together with dense overlap and occlusion, often lead to ambiguous instance boundaries. In addition, repeated downsampling and feature reconstruction can progressively erode fine structural details, hindering contour preservation and the separation of adjacent leaves. To address these interrelated challenges, we propose DMSHA-Net, a dynamic multi-scale feature fusion and hierarchical attention network for leaf instance segmentation. Its direction-aware Multi-Scale Feature Aggregation (MSFA) encoder captures complementary horizontal and vertical contextual information across multiple receptive-field scales, thereby improving the representation of diverse leaf morphologies. The Dense Feature Aggregation (DFA) decoder selectively integrates deep semantic and shallow structural features through stage-specific attention mechanisms, where self-attention at coarse resolutions models long-range dependencies, whereas lightweight channel recalibration at high resolutions refines local structures and boundaries. The Learnable Feature Fusion (LFF) module subsequently integrates multi-level semantic and boundary features using normalized learned weights. DMSHA-Net achieves Best Dice (BD) scores of 93.17%, 85.12%, and 90.91% on KOMATSUNA, MSU-PID, and the CVPPP-A1 subset, respectively, with corresponding foreground–background Dice (FBD) scores of 98.19%, 91.02%, and 98.24%. These results indicate that DMSHA-Net achieves competitive segmentation performance across the three datasets, particularly in scenarios involving pronounced scale variation and dense leaf overlap.
Brown Rot (Monilinia spp.) and Leaf Curl (Taphrina deformans) are principal fungal diseases affecting peach (Prunus persica L. Batsch) production, lacking validated AI-based diagnostic tools in tropical highland orchards. This study presents a compact Convolutional Neural Network (3658 trainable parameters), applied identically to fruit and leaf classification, integrated with a background-removal preprocessing pipeline and evaluated through stratified 5-fold cross-validation on 800 in-situ images from four orchards in Cómbita and Choachí, Colombia. The proposed architecture achieved mean accuracies of 82.8% (fruit) and 95.3% (leaf), with AUC values of 0.87 and 0.98, and a trained model footprint of approximately 100 KB, supporting storage- and bandwidth-efficient deployment. Benchmarked against ImageNet-pretrained MobileNetV3-Small and MobileNetV2 under an identical protocol, the proposed architecture matched or exceeded MobileNetV3-Small on leaf classification despite a 257-fold smaller parameter count, and achieved comparable or lower inference latency than both larger backbones. To our knowledge, this is the first validated system for simultaneous detection of both pathogens in Prunus persica under real field conditions, combining a compact, deployment-ready architecture with an ablation-verified preprocessing pipeline. The proposed model was deployed in the DurAPP web platform, giving peach growers in tropical highland regions a practical, low-footprint diagnostic tool suited to smallholder farming conditions.
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.
Abstract Agricultural pest and disease monitoring plays a vital role in ensuring crop productivity, reducing pesticide consumption, and promoting sustainable agricultural development. Although deep learning techniques have achieved remarkable success in plant health diagnosis, many existing models remain computationally intensive and are difficult to deploy on resource-constrained edge devices used in practical agricultural environments. To address these challenges, this study proposes a practical lightweight deep learning framework based on an improved ShuffleNetV2 architecture for real-time maize pest and disease recognition under complex field conditions.The proposed model incorporates the Ghost module to reduce redundant feature generation, the Efficient Channel Attention (ECA) mechanism to enhance feature representation, and the HardSwish activation function to improve nonlinear learning capability while maintaining computational efficiency. Extensive experiments were conducted on a maize pest and disease dataset containing multiple disease and pest categories collected under natural field conditions. Experimental results demonstrate that the proposed model achieves superior recognition accuracy while significantly reducing model parameters and computational complexity compared with several mainstream lightweight convolutional neural networks.The results show that the proposed method achieves an accuracy of 93.00%, a recall of 92.76%, and an F1-score of 92.42%, while maintaining extremely low computational cost (0.03 GFLOPs) and model size (1.16 MB). Furthermore, the proposed model was successfully deployed on a Raspberry Pi platform, demonstrating excellent real-time inference capability and low computational resource consumption. The framework is suitable for practical agricultural applications, including intelligent crop monitoring, UAV-assisted field inspection, and mobile diagnostic systems. By enabling rapid and accurate in-field identification of maize pests and diseases, the proposed approach supports timely crop protection decisions, reduces unnecessary pesticide application, and contributes to sustainable agriculture through practical edge-AI deployment.
Hyperspectral sensors have emerged as a promising approach in the study of plant diseases. The objective was to distinguish between healthy and inoculated seeds, and also to distinguish between genera of plant-pathogenic fungi in soybean seeds, using hyperspectral sensors combined with machine learning. The experimental design was a fully randomized factorial design with six algorithms (Simple Logistic Regression, Support Vector Machine, Artificial Neural Network, Random Forest, REPTree and J48 decision trees) and four phytopathogens (Sclerotinia sclerotiorum, Macrophomina phaseolina, Rhizoctonia solani, and Colletotrichum sp.) plus the control. Spectral analysis of the seeds was performed using a spectroradiometer (Ocean Optics) consisting of two sensors: NIR and Flame, covering the spectrum from 350 to 2500 nm. It was possible to distinguish between healthy and inoculated seeds, as well as identify the type of phytopathogen, based on each spectral signature. The Simple Logistic Regression and Support Vector Machine algorithms performed best. Hyperspectral sensors combined with machine learning constitute a promising tool for the detection of phytopathogens in seeds, enabling rapid and non-destructive analysis. This promising tool could serve as a complementary alternative to traditional diagnostic methods, which, although accurate, are time-consuming and rely on specialized labor.
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.
Rice panicle blast detection is an important task in plant disease phenotyping. Field-based detection remains challenging because infected spike regions are often small, sparse, elongated, and affected by overlapping panicles, complex backgrounds, and variable illumination. In this study, we propose RPB-YOLO11, a lightweight YOLO11-based detector designed for rice panicle blast detection. The model uses a Lightweight Ghost Backbone (LGB) to reduce redundant computation. It uses Anisotropic Axial Stripe Attention (A2SA) to represent elongated panicle structures. It also uses Focal Multi-Scale Attention (FMSA) for multi-scale feature refinement and Adaptive Geometric Shape IoU (AGS-IoU) for geometry-aware localization. The model was trained and evaluated on a rice panicle image dataset containing 1,055 training images, 69 validation images, and 169 test images. On the test set, RPB-YOLO11 achieved 76.09% mAP50, 45.44% mAP50-95, 73.98% precision, and 72.75% recall with 6.21 GFLOPs. Compared with the YOLO11n baseline, it improved mAP50, mAP50-95, precision, and recall by 2.73, 2.12, 1.64, and 2.11 percentage points, respectively. An Android-oriented inference application supports local image inference, detection visualization, class counting, and diseased-panicle incidence estimation. These results suggest that RPB-YOLO11 provides a practical approach for image-based rice panicle blast survey.
Reproduction assets foundThe paper links a public Hugging Face dataset used to establish the rice panicle blast detection dataset and a public GitHub release (data availability statement) containing the study's datasets/models.Dataset · publicsites, cultivars, growth stages, imaging conditions, and disease severities are still needed to evaluate generalization more fully.
Statements
Data availability statement
The 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://github.com/XuzheYang2Doc/RPB-YOLO11/releases/tag/rpb-yolo11 .
Author contributions
XY: Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review & editing. XZ: Conceptualization, Methodology, Visualization, Writing – original draft, Writing – review & editing. CX: Formal analysisOpen asset ↗XuzheYang2Doc/RPB-YOLO11 · rpb-yolo11lines:639-658Plant phenotyping relevance matchCrossref · checked 14 Sept 2026
Published1 Sept 2026Computers and Electronics in Agriculture
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 matchOpenAlex · checked 14 Sept 2026
UAV-based multi-view reconstruction is an important approach for high-precision, non-destructive 3D crop phenotyping. However, in greenhouse environments, UAV image acquisition is often restricted to sparse viewpoints because of UAV-induced airflow disturbances and the structural complexity of the greenhouse, which severely hinders accurate 3D phenotyping. To address this challenge, this study develops a task-driven phenotyping framework for constrained UAV viewpoints in greenhouse environments, integrating a vision-triggered flight planning strategy with an improved sparse-view 3DGS pipeline, termed SparseBerry-3DGS, for multi-view image acquisition and 3D phenotyping of greenhouse strawberries in GNSS-denied environments. Specifically, 3D Gaussian Splatting (3DGS) is improved by incorporating flow-guided initialization, depth supervision, and an adaptive pruning strategy, which effectively alleviate geometric collapse and floating artifacts under sparse-view conditions. Furthermore, sequential semantic masks generated by SAM2 are utilized to guide the segmentation of strawberry point clouds, thereby reducing background interference and segmentation errors. Experimental results show that the vision-triggered flight strategy enables stable capture of 16 surrounding images for each target fruit. Under sparse-view conditions, SparseBerry-3DGS improves reconstruction stability, with the average peak signal-to-noise ratio (PSNR) reaching 18.25 dB, corresponding to an 18% improvement. The SAM2-based segmentation module achieves high accuracy, with the mean intersection over union (mIoU) above 0.95. Geometric evaluation based on strawberry longitudinal diameter yielded an of 0.88, supporting the accuracy of fruit-scale geometric reconstruction. For weight estimation, five-fold cross-validation yielded an of 0.90 and an RMSE of 3.62 g, showing better predictive performance than models based on 2D projected area and standard 3DGS point clouds. This study provides a new approach for high-throughput, non-invasive digital crop phenotyping in greenhouse horticulture.
Automated plant disease classification from leaf images demands models that jointly achieve high accuracy and efficient training convergence. Standard deep learning approaches process images through a single feature pathway, limiting their ability to capture diverse visual manifestations such as color changes, texture patterns, and spatial distributions. This paper introduces DMCNNA-FBVL, a framework integrating two complementary innovations: (1) a Deep Multi-Component Neural Network Architecture (DMCNNA) employing three specialized pathways processing color-space statistics, texture descriptors, and raw image features before fusing them through SoftMax-weighted aggregation; and (2) Feedback-Based Validation Learning (FBVL), a training strategy that periodically blends validation-set gradients into weight updates to accelerate convergence. Experiments on the New Plant Diseases Dataset (87,000 images, 38 classes) show that DMCNNA-FBVL achieves 98.7% accuracy, 98.8% precision, 98.6% recall, and 98.7% F1-score, outperforming ResNet-50 by 3.2 percentage points ( 𝑝 < 0 . 0 0 1 ). The primary reported metrics are computed exclusively on an independent 10% hold-out test set, whereas the separate 10% validation partition is used during training for FBVL gradient blending and does not contribute to final test evaluation. Five-fold cross-validation confirms stability (98.7% ± 0.15%). Ablation experiments confirm additive gains, while FBVL reduces wall-clock training time by 15% through faster convergence.
Reproduction assets foundThe paper's plant disease classification experiments use the New Plant Diseases Dataset, which the authors state is publicly available on Kaggle. The authors' code and trained models are only promised 'upon acceptance' (no public URL), so they do not qualify as public assets.Dataset · publicral
monitoring systems.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have
appeared to influence the work reported in this paper.
Data availability
The New Plant Diseases Dataset used in this study is publicly available on Kaggle (https://www.kaggle.com/datasets/vipoooool/new-plant-diseases-dataset). Code and trained models will be made available upon
acceptance.
Funding
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit
sectors.
Acknowledgments
[Removed for double-blind review.]
CRediT authorship contribution statement
[Removed fOpen asset ↗Kaggle · new-plant-diseases-datasetpdf-raw-page:19 lines:1-59Plant phenotyping relevance matchEurope PMC · checked 14 Sept 2026
Abstract In Bangladesh, the potato (Solanum tuberosum L.) stands as an indispensable food and cash crop, deeply intertwined with national food security, rural livelihoods, and the broader agricultural economy. However, foliar diseases such as early blight and late blight frequently precipitate substantial yield losses and quality degradation when not identified and mitigated during the nascent stages of infection. Contemporary diagnostic paradigms remain predominantly manual and visual, relying heavily on agricultural professionals, which is often inefficient and inaccessible for remote farmers. While deep learning has demonstrated remarkable efficacy in automated plant disease recognition, existing methodologies frequently lack interpretability, disease severity quantification, and real-world field applicability. This paper introduces a comprehensive, interpretable deep learning-based framework utilizing EfficientNetV2-B0 for classifying potato leaf images into healthy, early blight, and late blight categories. By integrating Gradient-Weighted Class Activation Mapping (Grad-CAM), the model achieves high transparency, highlighting critical prediction regions. Furthermore, a severity assessment module estimates infection percentages, providing actionable treatment recommendations, ultimately enhancing agricultural decision-making.
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.
Southern corn leaf blight (SCLB) is caused by the fungal pathogen Bipolaris maydis (syn. Cochliobolus heterostrophus Drechsler) and is a common disease of fall crops of sweet corn. Phenotyping for SCLB resistance is performed through visual scoring, which is subjective and may limit genetic gain for this quantitative trait. As an alternative, we integrated computer vision (CV)-based phenotyping, genome-wide association studies (GWASs), and predictive breeding approaches to dissect the genetic basis of SCLB resistance. We utilized a sweet corn diversity panel with 693 genotypes, for which whole-genome resequencing produced a high-density single-nucleotide polymorphism (SNP) dataset. Broad-sense heritability for visual scoring ranged from 0.44 to 0.73, while CV-based phenotyping produced estimates ranging from 0.56 to 0.73 in multi-environment resistance trials conducted across 5 years and three locations. We performed GWAS using 16,755,210 SNPs and identified 41 associated SNPs. Genomic selection (GS) models on visual scoring phenotypes achieved moderate prediction accuracies under cross-validation of untested genotypes across characterized environments (0.22-0.47) and high prediction accuracies when predicting tested genotypes in uncharacterized environments (0.49-0.68). Using CV-based phenotypes for GS, we observed prediction accuracies of 0.45-0.47 under the untested genotypes in the characterized environments cross-validation scheme and 0.59-0.62 under the tested genotypes in the uncharacterized environments scheme. GS demonstrated reliability for ranking the individuals across a gradient of environments. These findings identify candidate loci and predictive breeding strategies to accelerate the development of resistant sweet corn cultivars.
Reproduction assets foundThe authors state that all datasets (phenotype data) and analysis code (CV phenotyping script, customized GAPIT script) are publicly available in their GitHub repository, which is listed in allowed_urls.Code · publiche images taken for each plot were saved in JPG format and analyzed using a CV method. Here, we refer to the CV method as a custom Python script written using the OpenCV library version 4.5.0, a set of tools for CV (Bradski, 2000 ). The Python script used for leaf CV image analysis is available in our public GitHub repository ( https://github.com/Resende‐Lab/SCLB‐Disease ).
FIGURE 1
Leaf imaging set up with QR‐coded plot IDs (bottom right) and color checker for computer vision phenotyping of southern corn leaf blight disease severity in sweet corn.
In the CT19 environment, a black cloth attached to a wooden board was used as the background. A wooden frame was used to clamp the leaves down toOpen asset ↗Resende‐Lab/SCLB‐Diseaselines:199-209Dataset · publicBLUP and BayesB model implemented in BGLR.
ACKNOWLEDGMENTS
This work was supported by the National Institute of Food and Agriculture USDA‐NIFA2018‐51181‐28419, USDA‐NIFA2019–05410, and USDA‐NIFA 2022–51181‐38333.
DATA AVAILABILITY STATEMENT
All the datasets and codes used in this study are available in the following repository: https://github.com/Resende‐Lab/SCLB‐Disease
REFERENCES
Amadeu , R. R.
,
Cellon , C.
,
Olmstead , J. W.
,
Garcia , A. A. F.
,
Resende , M. F. R.
, &
Muñoz , P. R.
( 2016 ).Open asset ↗Resende‐Lab/SCLB‐Diseaselines:566-596Plant phenotyping relevance matchOpenAlex · Crossref · checked 14 Sept 2026
Abstract Grapevine leaves have dorsiventral anatomy with distinct adaxial (upper) and abaxial (lower) surfaces. Although morphological descriptor lists and ampelographic literature provide information on both the upper and lower side characteristics, in practice, the color traits of the upper side have become the focus of scientific publications. This study introduces the practical application of the recently developed LeafLaminaMap software and the use of trichromatic color indices in grapevine characterization. We aimed to compare colorimetric information on the adaxial and abaxial leaf surfaces as well as to explore the potential of machine learning models in classification. Five statistical descriptors (mean, standard deviation, contrast, energy, and entropy) were calculated for 25 RGB-based color indices on both leaf surfaces of 120 samples collected from four grapevine cultivars (‘Chardonnay’, ‘Pinot noir’, ‘Sauvignon blanc’, and ‘Syrah’). Data was subjected to multivariate statistical analysis and machine learning classifiers. Results showed that the abaxial leaf surface had stronger cultivar-specific color signatures, supporting its suitability for cultivar discrimination. These findings suggest that RGB-based analysis of both adaxial and abaxial leaf surfaces has potential for grapevine cultivar discrimination, offering a new perspective for cost-efficient plant phenotyping.
Background Understanding the structure of plant seeds cultivated for human consumption and food manufacturing is vital to provide sustainable products as well as to investigate early growth stages. This includes structural variation between different plant species, varieties and cultivars depending on genetic setup, as well as structural modifications upon germination, aging and storing or seed treatment during processing. For plant seeds as multi-component biological materials, structural characterization must extend across multiple length scales, from molecular organization to cellular architecture. Results We apply scanning Small- and Wide-Angle X-ray Scattering (SWAXS) and X-ray Fluorescence (XRF) on yellow pea seeds to combine local structural information on the molecular scale with imaging of cellular structures on the micrometer scale, enabling a comprehensive analysis of hierarchical organization. To identify and characterize heterogeneous regions within the pea seeds, we implement a fitting-free, data-driven segmentation and analysis workflow based on machine learning tools. This approach allows for classification of structurally distinct domains and enables quantitative comparison across samples without relying on predefined models. Furthermore, we incorporate multi-modal analysis by combining structural imaging with complementary elemental information obtained from XRF. The integration of compositional and structural data provides deeper insight into structure-composition relationships. Conclusions This multi-scale, multi-modal approach opens new possibilities for investigating hierarchical structures and their development under diverse conditions and enables systematic comparison between different species or seeds at different developmental stages or exposed to different processing steps. The approach is broadly applicable to various kinds of samples and other hierarchically organized biological materials, which makes it a valuable technique for plant science as well as plant-based food science.
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 matchCrossref · checked 14 Sept 2026
Published1 Sept 2026Computers and Electronics in Agriculture
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.
To overcome YOLOv11’s limitations in complex field environments, this paper proposed SDD-YOLOv11n, a lightweight real-time detector for soybean diseases. The model reconstructed the backbone using GhostConv to minimize redundancy and integrates a C3k2_Star module to enhance small lesion detection against background noise. Additionally, a Detect Efficient (DE) head further compressed the architecture. Experimental results verified the model's efficiency, achieving a parameter count of 1.88 M and a weight size of 3.9 MB—reductions of 27.3% and 25% compared to YOLOv11n, respectively. Furthermore, the model maintained high detection performance with a Mean Average Precision (mAP50) of 75.6% and an F1-score of 69.3%, demonstrating its effectiveness in balancing architectural efficiency and accuracy in complex field environments.
Anton Anton · Supriadi Rustad · Guruh Fajar Shidik · Abdul Syukur
Pepper / chilliPotatoTomatoLeafClassificationVisualization / data managementDisease symptoms / severity
Early detection of plant leaf diseases is critical for minimizing crop losses and supporting precision agriculture. While Convolutional Neural Networks (CNNs) have demonstrated high accuracy in image-based diagnosis, conventional architectures may not optimally balance spatial localization and channel-wise feature refinement, particularly in multi-crop classification settings. This study proposes a redundancy-aware dual-attention architecture, termed ATSA-DenseNet, which integrates the spatial branch of the Convolutional Block Attention Module (CBAM-Spatial) with Efficient Channel Attention (ECA) within a DenseNet121 backbone. Unlike prior dual-attention frameworks that retain full CBAM and introduce channel-level redundancy, the proposed design isolates complementary spatial and channel mechanisms to improve representational efficiency without increasing computational complexity. The framework is evaluated on controlled multi-crop PlantVillage-derived datasets comprising tomato, potato, pepper, and maize. Across both 3-crop and 4-crop configuration, ATSA-DenseNet consistently outperforms baseline DenseNet121 and single-attention variants, achieving 99.94% accuracy and 0.9994 macro-F1 on the 4-crop setting while maintaining a lightweight footprint (6.96M parameters, 2.87G FLOPs). Grad-CAM visualizations indicate improved localization of disease-relevant regions compared to the baseline. While results are obtained under controlled imaging conditions, the findings demonstrate that redundancy-aware dual-attention enhances feature discrimination efficiency in multi-class agricultural classification tasks. Future work will extend validation to real-field datasets with natural variability.
Rice false smut is a major panicle disease that affects rice yield and grain quality and is an important target for resistance evaluation in breeding programs. Accurate field phenotyping is important for disease assessment and resistance screening, yet current assessment relies heavily on manual visual scoring and smut ball counting, which are laborious and subject to evaluator variation. In close-range single-panicle images, false smut balls are often small, dense, occluded, adhered, making automatic detection and severity grading difficult. To address these challenges, we developed a YOLOv12-RSLW detector by integrating RepGhost, SimAM, LSCD, and WIoU into YOLOv12. Detection boxes were then used to guide the Segment Anything Model for panicle and lesion mask extraction, allowing calculation of the lesion-to-panicle area ratio as a supplementary indicator for count-based severity grading. A total of 1911 original field images were collected. After augmentation, the dataset contained 5663 images, including 4531 training images, 566 validation images, and 566 test images. Detection performance was evaluated on the test set, while SAM segmentation was assessed using 80 manually annotated original images. YOLOv12-RSLW achieved 92.06% mAP@0.5, 91.46% precision, and 87.01% recall, with 3.45 M parameters and 6.0 GFLOPs. Compared with the baseline YOLOv12, mAP@0.5 and recall increased by 3.60 and 4.37 percentage points, respectively. Within the augmented dataset, 41.2% of samples initially assigned to Grade 1 and 28.1% of those assigned to Grade 2 met the area-ratio criteria for potential reassignment to higher grades. The framework provides a quantitative approach to rice false smut severity phenotyping and may support future resistance breeding after further validation.
Plant diseases greatly affect agricultural production, especially in developing countries, where prompt diagnosis can be quite challenging due to the limited availability of experts in real-time. Deep learning techniques for image analysis is gaining popularity and are increasingly considered an alternative to traditional manual inspection of plants. This research presents the evaluation of plant leaf disease detection system based on a convolutional neural network (CNN) optimized with different nature-inspired algorithms. The backbone model is based on the EfficientNet-B0 pretrained on ImageNet. Therefore, transfer learning is used to adapt the model to an updated PlantVillage dataset. Experiments have been conducted with multiple nature-inspired algorithms to improve generalisation and training efficiency of the prediction model. Different data preparation techniques have been carefully applied to the dataset, creating a unified approach to ensure consistency in the preprocessing pipeline for the training, validation, and testing phases. Our experiments indicate that application of the grey wolf optimizer (GWO) for tuning key hyperparameters of the model, including dropout, learning rates, and weight decay produced the best results, with an accuracy around 99.45%.
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.
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.
Background: Faba bean is an important legume crop valued for its nutritional and soil-enriching benefits, yet its productivity is severely affected by foliar diseases. Automated image-based detection using deep learning provides a rapid and reliable approach for early disease identification and improved crop management. Methods: This study developed a machine learning-based automated framework for multi-class classification of Faba bean leaf diseases using transfer learning with the VGG16 convolutional neural network. A dataset of 8,021 RGB images collected under natural field conditions was used, comprising four classes: healthy, rust, gall and chocolate spot. Images were resized to 224 × 224 pixels and normalized prior to training. The pretrained convolutional layers of VGG16 were frozen and a custom classification head with global average pooling and dropout regularization was added. Model performance was evaluated using classification metrics. Result: The proposed model achieved an overall classification accuracy of 92.34% and a macro-averaged F1-score of 0.9227 on the test dataset. Strong classification performance was observed across all disease categories, with particularly high predictive accuracy for healthy and rust classes. The findings demonstrate the effectiveness of transfer learning for plant disease detection and highlight its potential for scalable, automated crop health monitoring in precision 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.
Rice leaf diseases pose a significant threat to global food security by reducing crop productivity and causing substantial economic losses. The traditional diagnosis method is manual method, which is low in efficiency, subjective and not suitable for large-scale agricultural monitoring. Despite the advances in automated disease detection using deep learning methods like CNNs, GANs, and transfer learning models, these techniques remain highly computational, not very flexible, and struggle to perform well in different imaging scenarios. Considering these drawbacks, this paper introduces an Explainable Deep Q-Learning based CNN framework which employs CNN-based feature extraction and Deep Q-learning-based adaptive policy optimization for rice leaf disease classification. The proposed model continuously refines the classification actions through reward-based learning, which makes the model more robust in various agricultural imaging environments, in contrast to traditional supervised CNN models that have static classification decisions. The proposed model achieved 98.5% accuracy, 98.52% precision, 98.50% recall, and a 98.51% F1-score, outperforming existing CNN, GAN, reinforcement learning, and transformer-based methods. It also offers a high computational efficiency of 14.2 GFLOPs, 248 MB memory consumption, ~ 48 min of training time, and 6.8 ms inference time per image suitable for resource constrained applications in agriculture. The results demonstrate the effectiveness, scalability, and practical applicability of the proposed framework. The proposed framework performs well on benchmark datasets but more research in the deployment of the edge-devices under different real-world agricultural settings will be investigated in future work.
Reproduction assets foundThe paper trains its Deep Q-CNN rice leaf disease classifier on public Kaggle rice leaf image datasets, which are cited with explicit public URLs and qualify as paper-specific phenotyping image inputs. The authors' own derived data/analysis artifacts are only available upon request, so no authors' code or trained modelDataset · publicSoni Gautam. Rice Leaf Bacterial and Fungal Disease Dataset. Kaggle. Available:Open asset ↗Kagglepdf-page:24 lines:1-94Code / dataset availability confirmedCrossref · checked 15 Sept 2026
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 confirmedCrossref · checked 15 Sept 2026
Precision agriculture is becoming more and more of a challenge that requires the use of intelligent systems that are able to predict stress and prevent yield loss before it is too late. Traditional methods of agricultural surveillance are predominantly reactive with irrigation demands being based on thresholds or individual yield forecasts models that do not represent the intricate spatio-temporal interactions that exist between crop physiology, soil status, and environmental stresses. Besides, the majority of the current practices do not have an autonomous decision-making approach to preventive intervention which leads to inefficient use of water and slows down the response to stress. This paper suggests a cognitive UAV-assisted agro-surveillance system to predict yield vulnerability caused by crop stress and optimize adaptive irrigation with the help of spatio-temporal deep and reinforcement learning. The framework combines UAV-obtained RGB and multispectral and thermal imagery with measurements of soil sensors and meteorological data obtained with the Crop Health and Environmental Stress Dataset. A new GeoSpatio-TRiNet model is used to acquire long-range spatial relationship, time stress development, and diffusion of stresses across agricultural regions. The model predicts the vulnerability trajectories of the stress instead of the direct yield regression, and this allows early detection of yield risk. Such predictions serve to generate a cognitive environmental state of a Soft ActorCritic (SAC) reinforcement learning agent that autonomously computes zone-based irrigation behaviors to reduce the recurrence of stress at the minimum water usage cost. As shown by the results of the experiment, the proposed framework has a stress forecasting accuracy of 96.3% and performs much better than the traditional machine learning, CNN-based, and transformer-based baselines. The system also decreases the predicted yield vulnerability by 46.6 and enhances water-use efficiency by 41.1 as compared to irrigation strategies based on rules. The results confirm the usefulness of spatio-temporal intelligence with predictive control in terms of effectiveness, and the proposed framework is a scalable and sustainable solution to precision agriculture of the next generation.
Reproduction assets foundThe paper uses the public Kaggle Crop Health and Environmental Stress Dataset (UAV RGB/multispectral/thermal imagery plus soil/weather measurements and stress labels) as its phenotyping data source, and the authors provide an explicit public GitHub repository for the analysis code.Dataset · publicThe current research is based on the Crop Health and Environmental Stress Dataset, which is a publicly available
dataset on Kaggle, specially created to help perform a spatio-temporal analysis of crop health in response to changing
environmental and water-stress factors [26].Open asset ↗pdf-raw-page:10 lines:1-62Code · publicturn: Final zone-wise stress predictions 𝐶
𝑡
𝑧, Yield vulnerability trajectories 𝑉𝑡
𝑧, Optimal adaptive irrigation policy
𝜋∗
End Algorithm
Code availability:
The data used to support the findings of this study are included in the article.
Code availability:
The code used in this research work is available in the following link.
https://github.com/replyvenugopal/Cognitive-UAV-Driven-Agro-Surveillance
4. Result and Discussion
The architectural agro-surveillance solution, which is proposed to be executed by UAVs, is executed through a
modular and scalable software framework to guarantee reproducibility and extensibility. The experiments are all
performed in Python as a main programming languageOpen asset ↗github.com/replyvenugopal/Cognitive-UAV-Driven-Agro-Surveillancepdf-raw-page:24 lines:1-55Plant 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.
Existing conversational plant-phenotyping platforms are difficult for plant scientists to use and lack the reliability scientific research demands: failed analyses are reported as valid measurements rather than flagged as missing, statistical tests run without checking assumptions, predictions carry no uncertainty estimate, and specialised hardware limits accessibility. We present PhenoIntel, a lifecycle-aligned multi-agent web platform that turns the full machine-learning workflow into a reliable, user-friendly phenotyping system. Nine specialised agents divide the analysis into stages, from image collection through model selection, inference, and reporting, rather than handing the whole task to one AI manager. Independent checks separate these stages, and every agent reads from and writes to one shared, fixed-structure record, so an inconsistent output from one stage is caught before it reaches the next. Uncertainty is matched to each model family, conformal prediction, detection-confidence spread, or Monte Carlo Dropout, rather than applied uniformly, and quality thresholds adapt to crop and task instead of one global cutoff. When no suitable model exists, PhenoIntel can propose, validate, and integrate a new one on its own. The model repository spans ten trained models across five crops and four imaging modalities. Classification models reach Macro F1 of 0.78-0.996; object-detection models reach 0.96 mAP@50 with a 54% reduction in counting error over an unoptimised baseline; and a temporal model reaches held-out Macro F1 of 0.7050. PhenoIntel runs in a browser on standard hardware, requiring no GPU, and a 1,200-test automated suite confirms complete pipeline execution. Every result carries calibrated uncertainty, validated statistics, and FAIR-compliant provenance, a combination existing conversational phenotyping tools do not offer.
Reproduction assets found論文固有の解析コードとモデル資産を公開するGitHubリポジトリを本文中の根拠とともに確認しました。Code · publiccode, model checkpoints, and the 1,200-test automated suite referenced
throughout this paper are maintained in a version-controlled
repository, available at
https://github.com/Naren1704/PhenoIntel-InternshipOpen asset ↗Naren1704/PhenoIntel-Internshiplines:2047-2163Plant phenotyping relevance matchCrossref · checked 15 Sept 2026
Agricultural production faces significant annual losses due to plant diseases, with economic impacts exceeding 40 million dollars and contributing to acute hunger affecting over 281.6 million people in 2023. The timely and accurate identification of plant diseases through leaf image analysis is crucial to mitigate these losses and ensure global food security. This study proposes a lightweight Convolutional Neural Network model, inspired by the MobileNet architecture, designed to classify various plant leaf diseases efficiently. Leveraging a publicly available dataset, this research focuses on developing a model that balances high performance with computational efficiency, making it suitable for real-world applications in resource-constrained environments. The proposed model, named LDPNet, achieved an outstanding accuracy of 99.62%, alongside precision, recall, F1-Score, and AUC metrics of 99.16%, 99.08%, 99.11%, and 99.99%, respectively. A comprehensive comparative analysis was conducted against MobileNetV2 and a reference model from previous research, highlighting the superior performance of LDPNet in terms of both accuracy and efficiency. The results demonstrate that the proposed architecture not only maintains high classification performance but also significantly reduces the number of parameters, making it a practical and scalable solution for plant disease identification. This study contributes to the growing field of agricultural technology by providing a robust, lightweight, and efficient tool for early disease detection, with the potential to enhance crop management and reduce economic losses in agriculture.
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 matchOpenAlex · checked 5 Sept 2026
Introduction Leaf shape is a genetically determined crop phenotype, and its accurate classification underpins soybean germplasm assessment and genetic improvement. Manual classification is highly subjective and struggles to distinguish morphologically similar leaves, while mainstream supervised classification demands large labeled datasets and incurs high development costs. Efficient feature frameworks for soybean leaf categorization are still insufficient. Methods In this study, 581 biologically replicated terminal leaflets sampled from 194 soybean varieties were analyzed at the single-leaflet level using traditional morphological indices and novel leaf contour angular features. Unsupervised K-means clustering was used to classify soybean leaflet morphological phenotypes; t-SNE was applied exclusively for dimensional reduction visualization, while Welch's ANOVA combined with Games-Howell post-hoc tests was adopted to detect inter-cluster phenotypic differences. Clustering stability and external consistency against manual visual labeling were further quantified via Adjusted Rand Index to comprehensively verify the reliability of grouping outputs. Results The results revealed no significant difference in leaflet edge complexity (p = 0.41) between two manually divided leaf groups distinguished by overall leaf outline similarity; these two morphologically similar leaf clusters failed to be fully separated even though the first two principal components accounted for 90.2% of total variance. For K-means clustering, k = 3 achieved better overall performance with a Calinski–Harabasz (CH) index of 395.55, Davies–Bouldin (DB) index of 1.03, and silhouette coefficient (SC) of 0.38, compared with k = 4. Nevertheless, the angular feature attained an F-value of 951.62 in driving sample reallocation across clusters, serving as the core indicator for fine subdivision at k = 4. Under k = 4 clustering, all six morphological indices differed significantly among the four groups (p < 0.05). Additionally, the number of cross-clustered samples increased from 66 to 119 as k rose from 3 to 4, with 96.6% of cross-clustering attributed to the leaflet contour angular feature. Discussion This research provides a novel reference and technical support for the automated identification and fine classification of soybean leaf morphology.
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.
Barley ( Hordeum vulgare L.) is a major cereal crop whose soluble dietary fiber (SDF) offers significant health benefits, yet conventional SDF determination methods are time-consuming, labor-intensive, and destructive to samples. This study developed a rapid, non-destructive method for SDF quantification in barley using mid-infrared (MIR) spectroscopy combined with a Boruta-based partial least squares (PLS) hybrid approach. A total of 280 barley grain samples were subjected to Fourier-transform infrared (FTIR) spectral acquisition from 2000 to 650 cm -1 , with reference SDF values determined by the association of official analytical chemists (AOAC) 991.43 enzymatic-gravimetric method. The Boruta algorithm selected 165 informative wavenumbers out of 363, reducing the variable space by 54.5%. The developed Boruta-PLS model achieved excellent predictive performance with a coefficient of determination for the training set ( R 2 C ) of 0.9695 and for the test set ( R 2 P ) of 0.9601 and a root mean squared error for the training set (RMSE C ) of 0.7464%, and for the test set (RMSE P ) of 0.7340%, substantially outperforming the full-spectrum PLS model with an R 2 P of 0.7759 and an RMSE P of 1.7387%, as well as conventional wavelength selection methods including variable importance in projection (VIP), competitive adaptive reweighted sampling (CARS), and uninformative variable elimination (UVE). The selected wavenumbers were predominantly located in chemically relevant regions at 1000-1200 cm -1 and 1500-1700 cm -1 , confirming model interpretability. This Boruta-PLS approach provides a rapid, non-destructive, and cost-effective alternative for SDF quantification, with strong potential for high-throughput screening and real-time quality monitoring of barley samples.
In the process of agricultural intelligence, precise detection of plant organs serves as the foundation for core tasks such as crop phenotyping analysis and yield prediction. However, in complex field environments, small targets such as citrus flowers and shoots face challenges including scale variation, background interference, and dense occlusion, which severely impact detection accuracy. This study improves the YOLOv10 model by introducing the BAM (Bottleneck Attention Module) attention mechanism and GIoU (Generalized Intersection over Union) loss function, constructing a YOLOv10s-BAM-GIoU model suitable for citrus flower, fruit, and shoot recognition. The BAM attention mechanism enhances the model's feature extraction capability for small target organs under complex backgrounds through parallel channel and spatial attention branches; the GIoU loss function improves the localization accuracy of densely occluded targets by optimizing the geometric alignment between predicted and ground-truth boxes. Validation experiments were conducted on a self-constructed dataset. The experimental results show that the improved YOLOv10s achieves significant advantages in comprehensive detection accuracy, with an mAP50 of 89.1%, representing an improvement of 2.9%~9.5% over the original YOLOv10s and other comparative models. In fine-grained category detection, the model achieves mAP50 of 91.2%, 83.6%, and 92.5% for shoots, flowers, and fruits, respectively. Furthermore, while maintaining high detection accuracy, the model achieves a detection speed of 23.6 ms per frame, meeting real-time detection requirements. The research results demonstrate that the improved YOLOv10s model integrating the BAM attention mechanism and GIoU loss function achieves an optimal balance between accuracy and speed in citrus organ detection tasks, providing a preferred solution for field real-time detection systems.
Abstract Technological advances have expanded the adoption of digital technologies in agriculture, helping to reduce labour effort, increase profitability, improve crop efficiency and productivity, enhance product quality, mitigate environmental impacts, and promote human health. This context also extends to soybean farming, a sector of major economic importance in Brazil. Most importantly, Brazil has the global leadership in soybean production with biological nitrogen fixation (BNF) replacing chemical fertilisers. The research and evaluation of BNF is limited by manual counting of nodules, a time-consuming procedure. This study presents SoyNodules, designed for the automatic identification of soybean nodules, consisting of a dataset of images. The dataset includes 1,701 images acquired under controlled conditions: 1,662 images of soybean roots with nodules and 39 images of isolated nodules without roots. A total of 49,210 nodule instances are manually annotated with bounding boxes. SoyNodules was designed to promote reuse and interoperability in alignment with the FAIR principles (Findable, Accessible, Interoperable, Reusable) and to support the development, training, and evaluation of computer vision and deep learning methods for precision agriculture.
Reproduction assets foundThe paper is a data descriptor for SoyNodules, an annotated dataset of 1,701 soybean root/nodule images with 49,210 bounding-box annotations, publicly deposited on Zenodo with a DOI. The same repository also hosts the authors' annotation-format conversion script (AnyLabeling to Pascal VOC/COCO), per the Code AvailabilDataset · publicThe SoyNodules dataset, released as version 1.0, is publicly available on Zenodo [28]
at https://doi.org/10.5281/zenodo.22081914.Open asset ↗Zenodo · 10.5281/zenodo.22081914pdf-page:9 lines:1-43Plant phenotyping relevance matchCrossref · checked 13 Sept 2026
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.
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.
Emerged Technologies include imparted potentiality to generate adequate food to converge ultimatum of society. However, aspects like, climate change, plant disease, refuse at pollinators as well as others are demanding to farmers. The presence of diseases pulls up the maturation of their corresponding species. Plant disease recognition using image processing and machine learning (ML) with data collected from Internet of Thing (IoT) sensors has been receiving greater amount of interest in recent years, however timely disease detection remains a demanding issues. Despite early disease detection though reduces the risk of destruction to plants nevertheless the noise present in the sensor networks (i.e. Internet of Things) while collecting internet scrapped images compromise overall precision and accuracy. To address on this research gap, i.e., addressing noise with timely disease detection, in this work ML based plant disease detection method called, Intersection Histogram and Rotational Invariant Principal Component Regression (IH-RIPCR) is introduced. IH-RIPCR technique is dividing as pre-processing and feature extraction. First with the raw plant dataset obtained as input from Corn or Maize Leaf Disease Dataset, pre-processing is done employing Intersection Histogram based Contrast Enhancement model. Second with the obtained pre-processed contrast enhanced images is subjected as input to ML-based Rotational Invariance Principal Component Regression feature extraction model to extract relevant features pertaining to corn or maize leaf images for disease detection in an accurate and precise manner. Experiments are conducted with corn or maize plant leaf dataset with different existing methods to verify hypothesis potentiality of IH-RIPCR technique. Significance of technique is evaluated based on numerous parameters in terms of PSNR, processing time, precision, recall and accuracy respectively.
Sharma S, Lupo Y, Munoz J, Cochetel N, Nunez V, Gaspar A, Torres-Lomas E, Cantu D, Diaz-Garcia L.
GrapevineMultispectral / hyperspectralRoot
Adventitious root formation is critical for the cost-effective propagation of grapevine rootstocks, and poor rooting limits the adoption of new rootstocks derived from underutilized Vitis L. species. We evaluated 308 accessions representing 18 Vitis species over three growing seasons, scoring rooting at two developmental stages, callus-stage and post-transplant, together with root biomass, cutting weight, and a derived transplant- response index. Phenotypic variation was extensive within and among species, and species rankings depended on the trait considered: V. riparia , V. rupestris and V. californica ranked among the top five species for all four rooting traits, whereas V. arizonica and V. acerifolia rooted well at the callus stage but only intermediately after transplanting. Repeatability was moderate to high for root weight and callusstage rooting and lower for post-transplant rooting. Between-species differences accounted for most of the genetic variance in callus-stage rooting but little of that in cutting weight. After removing differences among species, accessions originating from wild sites with lower dry-season precipitation rooted better and produced more root biomass. Genome-wide association analysis of 3.4 million single-nucleotide polymorphisms identified 54 significant markers resolving into 18 independent loci across four traits. Candidate genes implicate auxin-linked cell proliferation, cell wall and lignin remodeling, and solute transport. Genomic and phenomic prediction achieved moderate accuracies across traits and seasons, including for previously unevaluated accessions, and combining spectral with genotypic data improved performance; accuracy was essentially flat between 5,000 and 50,000 markers. These results provide a framework for broadening the germplasm base of grapevine rootstock breeding. Plain Language Summary Grapevines are almost always grown as two plants joined together: a fruiting variety grafted onto a rootstock that supplies the root system. Nurseries build these plants from dormant cuttings, so a rootstock is only useful in practice if its cuttings root easily. Almost all commercial rootstocks descend from just three wild North American grape species, in part because cuttings of other species are believed to root poorly. We grew cuttings from 308 wild and cultivated grapevines representing 18 species over three years and measured how well each one rooted, first in the callusing room and again after the young plants were transplanted. Rooting ability varied a great deal, and several species outside the usual three rooted as well as the standards. Vines originally collected from places with drier summers tended to root best. We also located regions of the grape genome linked to rooting, and showed that the rooting ability of a vine can be predicted from its DNA or from light reflected by its leaves. Together, these results give breeders a way to screen a much wider range of wild grapevines for rootstock development before testing them in a nursery. Core Ideas Rooting ability varies widely across 18 Vitis species, well beyond the three used in rootstock breeding. Callus-stage and post-transplant rooting behave as genetically distinct stages of root formation. Accessions from collection sites with drier summers rooted better and produced more root biomass. GWAS resolved 18 loci implicating auxin signaling, cell wall remodeling, and solute transport. Genomic and phenomic prediction reached moderate accuracy and was insensitive to marker density.
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:314-336Code / dataset availability confirmedbioRxiv · checked 5 Sept 2026
Biomolecular condensates that persist through cell division must be reorganized and inherited, yet it remains unclear whether subtle defects before division are associated with later organelle or growth phenotypes. We examined the Chlamydomonas reinhardtii pyrenoid, a liquid-like condensate that concentrates ribulose-1,5-bisphosphate carboxylase/oxygenase (Rubisco), the photosynthetic CO2-fixing enzyme. As part of the algal CO2-concentrating mechanism, the pyrenoid raises CO2 availability around Rubisco. We generated an RBCS1-mGold Rubisco reporter and developed an unsupervised image-analysis pipeline combining a convolutional autoencoder and a one-class support vector machine. Using 4,905 wild-type single-cell images, augmented 22-fold to 107,910 image instances, we defined the range of normal pyrenoid morphology. A combined machine-learning and visual screen of approximately 21,000 insertional mutants yielded 17 pyrenoid integrity mutants (pim1-pim17). Differential reconstruction-error maps highlighted local deviations from the wild-type reference, including phenotypes difficult to classify by eye. Four-dimensional live imaging showed defects in matrix dispersal, partitioning of Rubisco-containing foci, or pyrenoid recondensation in multiple pim strains. Growth assays identified broad defects and phenotypes that became more apparent as CO2 supply decreased. Insertion-site mapping nominated candidate loci, including STT7, which encodes a chloroplast kinase best known for regulating photosynthetic light harvesting. Independent STT7-edited lines lacked detectable STT7 accumulation and showed pyrenoid-region reconstruction-error patterns, supporting an association between impaired STT7 function and altered pyrenoid morphology. These findings show that unsupervised image screening can extend forward genetics to subtle pyrenoid phenotypes accompanied by mitotic remodeling or growth defects.
Reproduction assets foundThe paper's custom machine-learning analysis scripts (CAE–OC-SVM pyrenoid screening pipeline) are explicitly stated to be publicly available on the authors' GitHub repository. Other data (microscopy files, anomaly scores) are only available upon request, so they do not qualify as public assets.Code · publicCustom scripts used for the machine-learning analyses are publicly available at https://github.com/Yamano-Lab/2025_Machine_Learning-based_screening .Open asset ↗Yamano-Lab/2025_Machine_Learning-based_screeninglines:103-119Plant phenotyping relevance matchCrossref · checked 15 Sept 2026
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.
Tomato leaf diseases must be identified early and accurately in order to reduce output loss and advance sustainable agriculture. Deep learning models have shown encouraging results in the identification of plant diseases, but their high processing requirements and inability to adjust to field-specific limitations sometimes make it difficult to implement them in real-world applications. For the purpose of accurately and efficiently classifying tomato leaf diseases, we present a convolutional neural network based on AlexNet that is lightweight and field-aware. Our proposed model is designed with less complexity than traditional architectures and is able to perform inference faster without sacrificing accuracy, making it suitable for real-time implementation in low resource agricultural environments. The algorithm was trained and tested on a dataset of 7704 augmented images of tomato leaves from 7 different disease categories. The Lightweight AlexNet achieved accuracy comparable to VGG variants and outperformed deeper models such as ResNet (96.81%) with lower parameter overhead. It was trained from scratch using TensorFlow & Keras on 100×100 pixel inputs, achieving a training accuracy of 99.15% and a validation accuracy of 99.74%. Moreover, an effective structure of the model enables the installation on edge devices, which offers a scalable precision farming solution. Our work helps to bridge the gap between deep learning research and real-world application in agriculture, allowing the development of real-field, resource-efficient disease detection systems.
Plant diseases can reduce crop quality and productivity, making early detection an important aspect of modern agriculture. Recent advances in deep learning, particularly Convolutional Neural Networks (CNN), have shown promising performance in image-based plant disease classification. This study proposes an explainable deep learning approach for multi-class plant disease classification using ResNet50 and EfficientNetB0 combined with Grad-CAM visualization. The experiments were conducted using the PlantVillage dataset consisting of 15 classes of healthy and diseased plant leaves.The research process included image preprocessing, data augmentation, transfer learning, model training, performance evaluation, and explainability analysis. The dataset was divided into training and validation sets with a ratio of 80:20. Model performance was evaluated using accuracy, loss, confusion matrix, precision, recall, and f1-score metrics. Experimental results showed that ResNet50 achieved the best performance with an accuracy of 92% and a validation loss of 0.19, outperforming EfficientNetB0 which obtained 76% accuracy and 0.82 validation loss. The classification report demonstrated that ResNet50 provided more stable and consistent predictions across most disease classes. Furthermore, Grad-CAM visualization successfully highlighted disease-relevant regions such as lesions, discoloration, and damaged leaf areas, improving the interpretability of the CNN model. The findings indicate that the combination of ResNet50 and Grad-CAM is effective for plant disease classification and provides better explainability for deep learning-based agricultural applications.
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.
Herbarium specimens are physical, verifiable records that form the basis of taxonomic knowledge and biodiversity research. Their large-scale digitization has produced extensive collections of high-resolution images and associated specimen metadata, creating conditions in which artificial intelligence (AI) can play an important role in plant taxonomy, collection management, and ecological research. Early AI applications have primarily focused on automated species identification based on individual specimen images. Although increasingly accurate, such approaches remain limited by their emphasis on single-specimen label prediction and by treating identification outputs as final analytical decisions. Recent methodological advances-including segmentation-based preprocessing, automated trait extraction, structured extraction of label data, detection of potentially misidentified specimens, and multimodal integration of visual, textual, and genetic information-extend AI applications beyond species identification toward broader analytical frameworks, encompassing taxonomic interpretation as well as ecological and biodiversity research. In these approaches, specimens are placed within a shared analytical space, and identification results are used to support comparisons across multiple specimens rather than being treated as final decisions for single individuals. This multi-specimen perspective enables quantitative examination of species boundaries, morphological variation, data inconsistencies, and taxonomic stability within curated collections. In this context, AI serves not as an ultimate decision-maker but as a decision-support tool embedded in expert-guided workflows and biodiversity knowledge infrastructures. These developments can be summarized as Integrative Taxonomic AI, an approach that employs learned morphospaces to interpret and refine taxonomic categories by integrating multimodal evidence and curated specimen data under expert guidance.
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.
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.
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 matchEurope PMC · checked 14 Sept 2026
Northeast China's Geng rice (Oryza sativa subsp. japonica) dominates the high-value rice markets in China due to its superior eating quality. However, current evaluation methods rely on either labor-intensive, subjective sensory protocols or low-accuracy, calibration-heavy near-infrared spectroscopy (NIRS), constraining breeding for high eating quality and market development. Here, we report a vision-based deep learning framework combining multi-population fine-tuning with industrial vision-language model (VLM) pre-training for Geng rice eating quality prediction. Trained on natural and recombinant inbred (RI) population datasets, our optimal model (Model 4) showed high cross-population stability. It achieved R 2 values of 0.98, 0.57, and 0.61 in a natural population validation set (35 cultivars), an independent DA-RI population (201 lines), and a randomly collected set (30 Northeast and 28 Southern cultivars), respectively, consistently outperforming the widely used Satake STA1B analyzer. Furthermore, our approach enabled the mapping of a novel, robust quantitative trait locus, qIVOE7, for Geng rice eating quality on Chromosome 7. Further analysis suggested that Model 4 appears to rely on the Hue dimension of the HSV color space for its predictions. This framework provides a high-accuracy prediction model and an easy-to-use tool for rice eating quality evaluation, accelerating high-quality rice breeding as well as the development of the high-quality rice market.
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.
Abstract A systematic review based on 174 Scopus-records of studies using YOLO-type one-stage detectors for detecting citrus fruits, their count, and yield estimation followed PRISMA guidelines 2020. The title/abstract-screening process, done in duplicate (κ=0.920) yielded 90 included study-records, followed by two further post-hoc exclusions. Each study in the 80 reporting on fruit-level detection showed an average precision of 89.6%, recall of 85.9%, and mAP@0.5 of 91.0%. However, coverage for any individual metric rarely exceeded half of the studies, and only 13% were able to report the more stringent mAP@0.5:0.95. Both YOLOv8 and YOLOv5 were each utilized as the backbone architecture by approximately 22.2% of the studies. From 2025, YOLOv11 has also been emerging. Half of all studies modified architectural components including attention modules, lightweight architectures, and variants of IoU loss functions. Original contributions are generally concentrated in downstream tracking, sensor fusion, and yield modeling rather than the detector itself. A custom-made seven-domain risk of bias tool was developed and utilized by two reviewers who arbitrated discrepancies (91.5%). Results showed that all but one of the reviewed studies had a high level of risk due to almost universal lack of statistical validation and limited dataset diversity; a sensitivity analysis excluding the most risky studies left the performance profiles nearly identical. We conclude that the field has converged around a common technical toolkit but continues to lack standardized benchmarks, multispectral data, and rigorous field-deployment validation.
Potato, one of the world's most important staple food crops, is highly susceptible to various foliar diseases that significantly affect its productivity and pose a serious threat to food security, thereby contributing to economic losses and impacting farmers’ income. Therefore, early and accurate detection is essential. Conventional detection methods rely primarily on manual observation, which is time-consuming and requires specialized personnel. In this study, five convolutional neural network (CNN) architectures were evaluated for the automatic classification of potato leaf diseases, including pretrained models (ResNet50, MobileNet, and VGG16) and models trained from scratch (AlexNet and LeNet-5). The dataset was constructed by integrating and selecting images from publicly available Kaggle repositories, resulting in a total of 6,691 images distributed across five classes: early blight, late blight, potato leafroll virus (PLRV), mosaic virus (PVY), and healthy leaves. Multiple experiments were conducted by varying hyperparameters such as batch size, optimizers, and the number of training epochs. The results show that VGG16 achieved the best performance, with an accuracy of 99.87%, outperforming the other architectures. Additionally, a mobile application based on the optimal model was developed for real-time detection. These findings demonstrate the potential of deep learning for intelligent and scalable agricultural diagnostic systems.
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.
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.
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.
Accurate quantification of plant disease severity is essential for evaluating host-pathogen interactions and assessing the effectiveness of disease management strategies. Traditional visual scoring methods and manual estimation of infected tissue are widely used but are often subjective and prone to observer bias. Digital image analysis offers an objective alternative by enabling automated identification and quantification of symptomatic plant tissues based on color and spatial characteristics. Here, we present a MATLAB-based image processing protocol for differentiating diseased and healthy plant tissue from digital leaf images. The workflow involves acquisition of standardized leaf images, conversion of RGB images into hue-saturation-value (HSV) color space, segmentation of diseased tissue using defined HSV thresholds, refinement of the segmented mask through morphological operations, and extraction of the whole leaf area. The protocol then calculates the diseased area and total leaf area in pixels and computes the percentage of infected tissue. The method uses MATLAB together with the Image Processing Toolbox and can be implemented using simple scripts. This protocol enables rapid and reproducible quantification of disease severity in plant leaves exhibiting visually distinct symptoms such as necrotic lesions or blight patches. By minimizing observer bias and providing quantitative measurements of infected area, the protocol offers a practical and reproducible approach for plant disease phenotyping and evaluation of disease management strategies across diverse plant-pathogen systems where diseased tissues can be clearly distinguished from healthy tissues under reasonably controlled imaging conditions. Key features • A reproducible MATLAB-based workflow for separating diseased and healthy plant tissue using color-space segmentation. • Applicable to plant diseases where symptomatic tissue contrasts clearly with healthy tissue (necrosis, blight lesions, rot patches). • Requires digital leaf images, MATLAB, and the MATLAB Image Processing Toolbox for image processing and disease quantification. • Enables rapid calculation of diseased leaf area and disease severity using automated pixel-based quantification.
Reproduction assets foundThe protocol explicitly deposits its authors' MATLAB image-processing workflow (HSV segmentation, mask refinement, pixel-based disease quantification) in a public GitHub repository with README instructions and example images.Code · publicGitHub repository containing the MATLAB source code, README file with installation and execution instructions, and representative example image(s): https://github.com/pankajborahmajuli-source/Leaf-Disease-Detection-MATLAB-Code/blob/main/README.mdOpen asset ↗Leaf-Disease-Detection-MATLAB-Codehtml-lines:112-148Plant phenotyping relevance matchCrossref · checked 15 Sept 2026
Published20 Aug 2026Eighteenth International Conference on Digital Image Processing (ICDIP 2026)Cited by 0 · OpenAlex ↗
Jiayuan Yang · Juntao Xiong · Mingyue Zhang · Baoxia Sun · Hongxing Peng
NeRF / 3D Gaussian SplattingFruit2D/3D reconstructionSegmentation
High-fidelity and lightweight 3D fruit models are a crucial foundation for phenotypic analysis and automated agricultural robotic operations. However, in complex agricultural scenarios characterized by varying illumination and foliage occlusion, existing 3D reconstruction methods struggle to balance reconstruction accuracy and model size, often generating massive redundant background primitives. To address this challenge, this paper proposes a novel framework for lightweight and high-fidelity 3DGS fruit reconstruction via geometric-semantic joint constraints. Specifically, the method first integrates depth priors and semantic information through a Depth-Guided Semantic Segmentation module to extract accurate target fruit masks. Next, it eliminates background noise points from the initial point cloud using a multi-view Reprojection Consistency Voting mechanism. Simultaneously, a Stochastic Background Regularized Hybrid Loss is introduced during the 3DGS training phase to decouple density and color optimization, thereby suppressing the regeneration of background Gaussian primitives. Experimental results on a multi-category fruit dataset demonstrate that while maintaining a high novel view synthesis quality (PSNR of 31.87 dB), our proposed method reduces the average model size from 230.42 MB to 62.66 MB (a 72.8% reduction), achieving robust, high-fidelity, and lightweight 3D fruit reconstruction.
Cell polarity and tip growth rely on the dynamic spatial organization of signaling and structural components. Quantitative characterization of these spatiotemporal dynamics is critical for understanding polarized cell growth, yet manual quantification is labor-intensive and existing computational tools often lack the flexibility and robustness needed to analyze molecular and structural dynamics in tip-growing cells. Tip Quantification (TipQuant) identifies the cell apex by detecting the site of maximum expansion and automatically quantifies fluorescence distribution along the plasma membrane and within the apical cytoplasm from live-cell imaging data, enabling analysis of the spatiotemporal dynamics of molecular and structural components in tip-growing cells. TipQuant accurately identified cell apices and quantified the spatiotemporal behavior of fluorescently labeled proteins and cellular structures in Arabidopsis thaliana pollen tubes and Fusarium graminearum hyphae, reproducing manual measurements while reducing user bias and improving efficiency, consistency, and analytical flexibility. The tool also revealed a strong positive correlation between rho-like GTPase from plants activity and apical Ca 2+ influx in Arabidopsis pollen tubes, demonstrating its utility for analyzing dynamic cellular processes. TipQuant is a robust analytical tool for quantifying spatiotemporal dynamics in tip-growing cells, providing a flexible alternative to manual image analysis and enabling studies of the molecular mechanisms underlying polarized growth.
Introduction In a complex orchard environment, canopy apples are obscured by various factors, making it hard for apple harvesting robots to accurately determine which apples can be directly harvested. Furthermore, the complex obstruction leads to difficulties in identifying keypoints on the apples and caculating the axis direction, directly affecting the robot's determination of grasping positions. Methods To solve these issues, a two-dimensional (2D) axis estimation method for pickable canopy apples based on a YOLO cascade network was proposed. Firstly, the study introduced SPDConv for lossless downsampling and adopts dynamic upsampling to improve segmentation boundary accuracy, constructing an instance segmentation network named the YOLO-SD model to select pickable apples according to different occlusion conditions and growth states. Secondly, the geometric center of the mask image was located using its minimum enclosing circle, and the region of interest was extracted through morphological dilation. Then, by integrating the RFAConv, SCSA attention, and MBConv modules, a keypoint detection network YOLO-RSM was constructed to extract keypoints of pickable apples. Finally, a 2D axis construction strategy was proposed, which adaptively constructs the growth axis based on the visibility of keypoints. Results Experimental results show that the overall average accuracy mAP50 of the YOLO-SD model for apple segmentation reached 95.2%, and the parameter quantity was reduced to 2.47 M. The average accuracy of the YOLO-RSM model for keypoint detection has reached 90.3%, which is 2.4%, 2.6%, and 4.7% higher than that of the YOLOv8n, YOLO11n, and YOLO12n models respectively. The 2D axis estimation algorithm has an average axis angular error of 7.23° ± 16.73°, and an axis estimation accuracy of 92.68%. Discussion The proposed method can achieve high-precision canopy apple segmentation, keypoint detection, and 2D axis estimation, thus offering technical support for the picking operations of apple harvesting robots.
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.
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.
Abstract Satellite-based prediction of grain protein concentration (GPC) in wheat typically composites spectral observations over fixed calendar windows, implicitly assuming phenological synchrony across fields. We present a systematic evaluation of whether aligning multi-source remote sensing time series to field-specific, spectral-peak-relative windows improves field-level GPC prediction, for a quality trait whose physiology, senescence-linked nitrogen remobilization, contrasts with the season-integrating behavior of yield. Integrating Sentinel-2 imagery (31 vegetation indices, 10 spectral bands), ERA5-Land reanalysis, gSSURGO soil properties, and USGS 3DEP topography across 228 commercial winter wheat fields in western Kansas (2024–2025), we compared six temporal strategies (peakrelative vs. calendar × monthly, biweekly, growth-stage) using three ensemble tree models under nested cross-validation with Boruta feature selection. A single 30-day post-peak window (peak + [16,45] days) was the top-performing and most consistently selected window, chosen in 4 of 5 outer folds, reproducing prior accuracy under random cross-validation (R2 ≈ 0.28); though its advantage over the best calendar window was not statistically significant (paired bootstrap p = 0.08). Under leave-county spatial cross-validation, however, this skill did not transfer across counties (Sentinel-2–only R2 ≈ 0.01; per-county median R 2 = −0.23), indicating the satellite signal supports within-region interpolation but not spatial extrapolation to unseen counties; ablation shows that neither the spectral nor the static features transfer across counties on their own, and the residual crosscounty skill emerges only from their combination. A near-real-time application at ∼3 weeks before harvest retains most within-region skill at a modest accuracy cost. The results delineate where spectral-peak-relative alignment helps, concentrating a senescence-linked signal within region, and where it does not, providing an honest operational baseline for satellite-based grain-quality monitoring.
Reproduction assets foundThe preprint explicitly releases the authors' analysis code (data-acquisition pipeline, feature engineering, cross-validation/modeling, figure scripts) at a public GitHub repository, and a de-identified field-level GPC dataset released alongside the code repository. Both are paper-specific, public, and actionable. The Code · publicthe figure-generation scripts is available at https://github.com/Ciampitti-Lab/Open asset ↗Ciampitti-Labpdf-page:48 lines:1-55Dataset · publica de-identified version of the dataset is released alongside the code repositoryOpen asset ↗pdf-page:48 lines:1-55Code / dataset availability confirmedEurope PMC · checked 5 Sept 2026
Introduction Accurate classification of unsound wheat kernels is important for automated grain quality assessment, but improved recognition performance often comes at the cost of increased model complexity. Methods This study presents ECA-ModNet, a parameter-efficient convolutional network derived from EfficientNetV2-S. The architecture replaces two early-stage Fused-MBConv blocks with Mod-FusedMBConv blocks to introduce input-dependent local contextual modulation and replaces the squeeze-and-excitation modules in later stages with efficient channel attention to model local cross-channel interactions using fewer attention-related parameters. Experiments were conducted on the seven-class G600 wheat subset of the GrainSpace dataset. Results Across three independent runs, ECA-ModNet achieved an accuracy of 90.17 ± 0.21% and a macro-F1 score of 90.23 ± 0.21%, improving upon EfficientNetV2-S by 3.80 and 3.84 percentage points, respectively. The parameter count decreased from 20.19M to 16.49M, while FLOPs increased marginally from 2.90G to 2.95G. ECA-ModNet achieved accuracy statistically comparable to that of ConvNeXt-Tiny and InceptionNeXt-T while using substantially fewer parameters, and obtained 3.03-4.55 percentage points higher mean accuracy than six lightweight baselines. Discussion Ablation experiments identified two Stage 1 Mod-FusedMBConv blocks with a 3×3 context kernel as the configuration with the highest mean accuracy among those evaluated. These results indicate that ECA-ModNet offers a favorable accuracy-parameter trade-off for image-based classification of unsound wheat kernels.
Reproduction assets foundThe paper analyzes the public GrainSpace dataset (G600 seven-class unsound wheat kernel subset) and provides an explicit data availability statement with a public GitHub URL. No author analysis code or trained model deposit is stated.Dataset · publicPublicly available datasets were analyzed in this study. This data can be found here: https://github.com/hellodfan/GrainSpace .Open asset ↗hellodfan/GrainSpacelines:1035-1076Plant phenotyping relevance matchCrossref · checked 5 Sept 2026
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.
Abstract Understanding the allometric relationships between leaf area and other plant traits is essential for non-destructive growth monitoring and efficient crop management. However, no comprehensive study has yet modeled leaf area in quinoa ( Chenopodium quinoa Willd.) using simple morphological traits across different sowing dates. This study aimed to quantify allometric relationships between leaf area and plant height, leaf dry weight, stem dry weight, panicle dry weight, and total dry matter, and to evaluate whether these relationships are modified by sowing date. A two-year field experiment was conducted with 12 sowing dates under a randomized complete block design with three replications. Leaf area index (LAI) dynamics were described using a logistic model, and allometric relationships were fitted using power-law equations. The results showed that LAI followed a logistic trend across all sowing dates, with maximum values ranging from 2.7 to 7.9. Plant height provided the most reliable prediction of leaf area (R² = 0.83, b = 1.2), followed by leaf dry weight (R² = 0.72, b = 0.97). The allometric coefficients for stem dry weight (b = 1.54, R² = 0.74) and panicle dry weight (b = 1.95, R² = 0.71) showed greater variability. A striking finding was the exceptionally high allometric coefficient (b = 4.95) recorded on May 6 of the second year, indicating a pronounced shift in resource allocation toward leaf area expansion. Total dry matter was a weak predictor (R² = 0.54), likely due to leaf fall during the growing season. The hypothesis that sowing date modifies allometric relationships was confirmed, as evidenced by considerable variation in allometric coefficients across sowing dates. This study provides, for the first time, a comprehensive set of allometric models for quinoa across multiple sowing dates. Plant height and leaf dry weight are recommended as simple, rapid, and non-destructive indicators for leaf area estimation, facilitating improved crop monitoring and management under diverse environmental conditions.
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:266-273Code / dataset availability confirmedOpenAlex · Crossref · checked 15 Sept 2026
Abstract Heat stress causes ultrastructural damage in pollen grains, leading to reduced pollen germination, pollen size and shortened pollen tube length, ultimately lowering seed set and yield. This study presents a high-throughput phenotyping framework that integrates controlled-environment pollen germination assays with deep learning–based object detection for rapid, accurate, and scalable evaluation of reproductive heat tolerance in soybean breeding programs. Sixteen soybean genotypes were grown under controlled environments at optimal (28/18°C; day/night) and high temperature (38/28°C; day/night) regimes during flowering. In vitro pollen germination was quantified using six YOLO (You Only Look Once) object-detection architectures (YOLOv7–YOLOv12) to identify the best-performing model for automated analysis. Among the tested object-detection architectures, YOLOv9 achieved the best overall performance for detecting germinated and non-germinated pollen grains in complex images. High temperature significantly reduced mean pollen germination from an average of 40% under optimal conditions to an average of 21% under heat stress (P < 0.05), with a significant genotype × growth temperature interaction. Invitro incubation temperatures ranging from 10 °C to 45 °C produced a clear thermal response; however, no significant genotype × incubation temperature interaction was detected within either growth temperature regime. Although photosynthetic and physiological traits were measured exploring their relationship with pollen germination, their transient and complex response limited their reliability for predicting reproductive performance. The automated pipeline substantially reduced the time required to evaluate pollen germination. The pipeline processed nearly 5,000 images in approximately one hour, substantially increasing throughput and reducing reliance on manual counting. The findings demonstrate that pollen germination is a promising proxy trait for screening reproductive heat tolerance in soybean. Combining controlled environment phenotyping with YOLO-based object detection enabled efficient, accurate, and scalable pollen analysis, and represents the central methodological advance of this study. YOLOv9 performed best among the tested architectures, although discrepancies from manual counts in some images indicate that additional validation is needed. The weak associations with vegetative physiological traits further support the value of direct pollen-based phenotyping.
Reproduction assets foundThe authors state that all data supporting the study, including annotated pollen germination images, computational and statistical codes, and analysis tools, were deposited in Zenodo with a public DOI. This is a paper-specific, publicly actionable asset. LabelMe and Ultralytics YOLO are generic third-party tools, not作者Dataset · publicCommission.
Data availability
All data supporting the findings of this study, including annotated images, computational and
statistical codes, and analysis tools, have been deposited in the Zenodo data repository. Additional
data will be made available upon reasonable request following acceptance of the manuscript.
Repository: https://doi.org/10.5281/zenodo.21685593
Ethics approval and consent to participate
Not applicable
Consent for publication
Not applicable
Competing Interests
Authors declared no competing interests
References
1. FAOSTAT: Crops and livestock products: soybean production data. https://www.fao.org/faostat/
(2022). Accessed 15 Feb 2026.
2. Patel D, Franklin KA. TemperaturOpen asset ↗Zenodo · 10.5281/zenodo.21685593pdf-raw-page:28 lines:1-34Code / dataset availability confirmedOpenAlex · Europe PMC · checked 15 Sept 2026
Boll-opening concentration is critical for mechanical cotton harvesting, yet it is still assessed mainly by manual records and single time-point indicators that miss temporal dynamics. To bridge the lack of a unified workflow linking vision foundation models, multi-temporal boll-opening monitoring, and harvest decision-making, we developed a cross-scale UAV high-throughput phenotyping framework centered on DINO-BollGX. DINO-BollGX couples a DINO v3 backbone, a RetinaNet detection head, and an adaptive refinement-and-suppression module for robust open-boll detection under complex field conditions. Using multi-temporal UAV imagery collected over two years for 383 cultivars, we reconstructed plot-scale time series of open-boll counts, derived dynamic features describing progression and intensity changes, and proposed a Cotton Boll-Opening Temporal Stability Index (CTSI) to quantify boll-opening rhythm and concentration; CTSI was further integrated with a time-based risk function to generate harvest decision curves. Under unified data and training settings, DINO-BollGX achieved precision = 0.91, F1 = 0.88, and AP@0.50 = 0.80, and provided accurate boll-count estimation (R 2 =0.98; MAE=3.10), outperforming YOLOv11, YOLOv12, YOLOv13, and RT-DETR. On an independent cross-year dataset acquired at 5 m altitude, it obtained precision = 0.98 and F1 = 0.87. An internal consistency analysis showed that CTSI had the expected negative association with Window_days (r = −0.83) and positive associations with Max_count (r = 0.78) and the boll-opening efficiency index (r = 0.92), reflecting the co-occurrence of temporal compactness and main-phase opening intensity in the cultivar population. CTSI ranged from −2.72 to 4.69 across cultivars, enabling identification of highly synchronized boll-opening. Harvest decision curves indicated that the relative net income index peaked at day 67 after the first observation and a compact optimal harvest window near the end of monitoring; on a fixed harvest date, Kuche 130292 (CTSI=4.69) produced 486 open bolls versus 182 for Xinluzao 36 (CTSI=0.53) and 90 for Andizhan-60 (CTSI=-2.72). Overall, the framework integrates dynamic boll-opening phenotyping with harvest timing optimization, supporting scalable cultivar screening and mechanization-ready deployment, with potential extension to harvest decision scenarios in other crops.
Reproduction assets foundThe paper publicly releases its cotton boll-opening UAV image dataset (3638 patches, 94,774 YOLO-format bounding-box annotations) on GitHub, directly supporting the paper's phenotyping analysis. No author analysis code or trained model checkpoints are explicitly deposited.Dataset · publicentary information for evaluating cross-scale detection performance and characterizing macroscopic spatial patterns. The 5 m imagery acquired on 18 Sept 2024 is used exclusively for cross-year generalization assessment. All cropped images and the corresponding YOLO-format annotation files have been publicly released on GitHub ( https://github.com/mianchen0529/cotton-boll-dataset/tree/main ) to facilitate further research on cotton phenotyping, agricultural remote sensing, and intelligent analytics.
2.3.
Model construction
2.3.1.
Overall architecture of the DINO-BollGX network
The proposed DINO-BollGX network consists of four stages: image preprocessing, feature extraction, object prediction,Open asset ↗mianchen0529/cotton-boll-datasetlines:63-74Plant phenotyping relevance matchOpenAlex · arXiv · checked 15 Sept 2026
Recent 3D foundation models provide powerful feature representations for point cloud learning by controlling spatial granularity. However, relying on a fixed spatial granularity severely limits generalization in applications like plant phenotyping, where organ morphology and size vary substantially across species and growth stages. To address this, we propose AGS-PlantSeg, a few-shot 3D plant organ segmentation method that leverages the frozen Utonia (arXiv:2603.03283) foundation model combined with Adaptive Granularity Selection. By dynamically selecting the best granularity levels for each specific plant model, our method extracts optimized geometric features for a lightweight MLP segmentation head. Extensive experiments across PLANesT-3D (arXiv:2407.21150), Pheno4D , and Crops3D demonstrate that AGS-PlantSeg significantly improves cross-species generalization, achieving 88.9% average mIoU performance and outperforming fixed-granularity baselines by 2.5 mIoU points. Despite requiring minimal annotated data, our approach is highly competitive with fully supervised, plant-specific architectures.
Image-based plant disease identification is essential for advancing smart agriculture, particularly for staple crops such as rice and economically significant crops such as tea, which are cultivated under complex environmental conditions. Furthermore, these crop groups present distinct challenges: rice leaf disease data typically exhibits clear pathological structures but necessitates large-scale deployment on resource-limited devices, whereas tea leaf disease data is complicated by variable lighting, diverse backgrounds, and high biodiversity. Although current deep learning models achieve high accuracy, they predominantly utilize deep convolutional neural network (CNN) architectures with millions of parameters, which hinders practical deployment on edge devices and increases the risk of overfitting when field data is scarce. In response, this study introduces DisQuan, a hybrid architecture that integrates classical deep learning with quantum machine learning (QML) to balance accuracy and resource efficiency. In particular, DisQuan combines the lightweight DisNet feature-extraction network with a variablequantum neural network to compress and refine feature representations in quantum space, yielding a model with only 0.09 million parameters. Experimental results on rice and tea leaf disease datasets indicate that DisQuan achieves the highest accuracy on the rice dataset and performance comparable to deep CNN models with significantly more parameters on the tea dataset, while maintaining a compact and stable structure. Overall, these findings suggest that DisQuan provides a practical compromise between performance and model complexity, and highlight the potential of quantum-classical hybrid architectures for plant disease detection in real-world agricultural settings and on resource-constrained devices.
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.
Dorado-Betancourt H, van Eeuwijk F, van Etten J, Ramirez-Villegas J, de Sousa K, Daudi H, Ndegwa M, Mendes T, van Heerwaarden J.
MaizePeanut / groundnutSweet potatoField / plot
Key message Our scalable two-step method estimates genotypic performance and genetic parameters from ranking data, producing reliable results comparable to quantitative analyses, enabling the integration of ranking data into breeding pipelines. Plant breeding research has chiefly relied on on-station experiments to evaluate varietal performance. Nevertheless, these trials often fail to represent on-farm growing conditions and farmers' preferences, potentially leading to poorly defined breeding targets. Recent work has demonstrated the potential of using on-farm verification trials combined with ranking data to support farmers in evaluating varieties while providing information that is representative of farmers' needs. Despite this potential, scalable methods for quantifying genetic differences and assessing the strength of the genetic signal in such trials remain limited. Here, we present a two-step procedure for analyzing trials based on ranking data, allowing the estimation of genetic parameters. The approach follows a common strategy in quantitative genetics, in which parameters are estimated from tables of genotypic means and their variances. In our framework, these estimates are obtained from Thurstonian and/or Plackett-Luce models, which treat rankings as observations of an underlying continuous trait associated with genotypic performance. Using simulated data, we showed that genotypic mean estimates derived from ranking analyses are linearly related to those obtained from quantitative trait analyses and that their variances adequately capture estimation uncertainty. We further demonstrated that incorporating these estimates and their variances into a second-step mixed-effects model yields accurate estimates of variance components. Analyses of groundnut, maize, and sweetpotato datasets confirmed the applicability of the approach and showed that ranking data can provide reliable estimates of genetic parameters. We argue that this framework can be scaled to obtain genotypic performance estimates from multi-trial on-farm data.
Reproduction assets foundThe paper's Data availability statement provides public access to the observed groundnut and sweetpotato ranking/trial datasets (Zenodo 17112492), the authors' R functions and simulation workflow (GitHub hdorado/tricot-ranking-analysis, archived Zenodo 17942919), and supplementary material with methods and figures (ZenDataset · publicThe observed data for groundnut and sweetpotato used in this study are publicly available and can be accessed at: Global multi-crop agricultural trial data supported by citizen science, Zenodo [ https://doi.org/10.5281/zenodo.17112492 ]Open asset ↗Zenodo · 10.5281/zenodo.17112492lines:205-225Code · publicThe R functions and simulation workflow used in this study are publicly available at: - Source code available from: [ https://github.com/hdorado/tricot-ranking-analysis ]Open asset ↗GitHub · hdorado/tricot-ranking-analysislines:205-225Code · public- Archived software available from: [ https://doi.org/10.5281/zenodo.17942919 ] - License: [MIT License]Open asset ↗Zenodo · 10.5281/zenodo.17942919lines:205-225Plant phenotyping relevance matchCrossref · checked 14 Sept 2026
Published17 Aug 2026NOUN Interdisciplinary Journal of Computing, E-Learning & Application (NOUN-IJCEA)Cited by 0 · OpenAlex ↗
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.
Rice seed vigor is a key determinant of germination performance and final crop yield, making its rapid and non-destructive assessment essential for seed quality evaluation. Conventional vigor detection methods are often destructive, labor-intensive, and time-consuming. Hyperspectral imaging provides a promising non-destructive alternative, but hyperspectral data are typically high-dimensional, redundant, and susceptible to noise and scattering interference. Moreover, existing models still have limited ability to discriminate subtle spectral differences among seed vigor levels. To address these challenges, this study proposes a gated recurrent unit (GRU)-guided closed-loop CNN-Transformer network (GCT-BCLN) for accurate, non-destructive identification of rice seed vigor. The model establishes bidirectional information flow between CNN and Transformer via the GRU, enabling dynamic and synergistic optimization of local spectral features and global spectral representations. In addition, a combined preprocessing strategy integrating adaptive iteratively reweighted penalized least squares (AirPLS), Savitzky-Golay (SG) smoothing, and multiplicative scatter correction (MSC) was adopted to improve spectral quality. Experimental results showed that GCT-BCLN achieved a test accuracy of 0.9795 for hybrid indica rice, outperforming the CNN-Transformer fusion model by 1.37%. The model also achieved accuracies of 0.9793 and 0.9758 on conventional japonica rice and glutinous japonica rice, respectively, showing consistent performance across the three evaluated variety-specific datasets under the controlled experimental protocol. These results support the feasibility of GCT-BCLN for laboratory-scale, non-destructive discrimination of aging-induced rice seed categories under controlled conditions, while practical application requires further external validation.
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.
Field / plotLeafSegmentationDisease symptoms / severity
Camellia oleifera leaf disease segmentation under natural field conditions is important for precision plant protection but remains challenging because lesions often show small target areas, blurred boundaries, uneven illumination, complex backgrounds, and coexisting symptoms. To address these problems, this paper proposes BFMambaNet, a Boundary-Frequency-guided Global Semantic Mamba Network for fine-grained disease segmentation. The model adopts an encoder-decoder framework and introduces a Global Semantic Mamba-based spatial selective feature modeling block to capture long-range lesion context and reduce semantic confusion. A gated wavelet spatial enhancement block is further designed to strengthen high-frequency boundary details while suppressing noisy responses. During training, boundary-frequency auxiliary supervision guides contour localization and pathological texture recovery without additional manual boundary labels. A reinforcement-learning-guided adaptive loss controller adjusts class-wise reweighting factors and loss-component weights according to the training state, improving optimization stability. A pixel-level dataset containing 1400 images and seven disease categories was constructed for evaluation. Experimental results show that BFMambaNet achieves 92.39% Precision, 91.43% Recall, 91.26% Dice, and 85.46% mIoU, outperforming representative CNN-based, Transformer-based, and Mamba-based models. Evaluations on environmental subsets confirm superior robustness, outperforming VMamba by 3.70% mIoU under uneven illumination, 3.55% mIoU under complex backgrounds, and 5.10% mIoU under coexisting symptoms. Cross-dataset validation on Apple leaf diseases further proves its generalization with 3.39% mIoU and 3.84% Dice improvements over U-Mamba, while maintaining a competitive inference speed of 30 FPS. Qualitative results also show clearer boundaries, fewer missed small lesions, and more stable predictions in complex field scenarios.
Introduction Applying general-purpose vision-language models (VLMs) to crop disease diagnosis presents three critical bottlenecks: reliance on large-scale annotated data, the high computational cost of full finetuning, and existing adaptation methods designed mainly for discriminative classification without sufficient visual-linguistic interaction for generative diagnosis. Methods We propose BLAP, an adaptive multi-scale visual prompt fine-tuning framework built upon BLIP-2. BLAP introduces an adaptive visual prompt fusion module (APFM) with learnable prompt vectors and a gating mechanism, together with a multi-scale pyramid feature fusion module (PFM). All BLIP-2 backbone parameters are frozen, and only 0.11% of the model parameters are optimized. Results On a few-shot dataset comprising 990 images from 11 crops and 33 disease categories, BLAP achieved 92.78% recognition accuracy, outperforming the BLIP-2+LoRA baseline by 21.67 percentage points. BLEU-4 and ROUGE-L scores reached 0.6507 and 0.7184, respectively, while inference latency increased by only 2.15%. Discussion BLAP provides a lightweight solution that balances accuracy, efficiency, and interpretability for crop disease diagnosis in resource-constrained settings. The proposed dynamic prompt fusion and multiscale pyramid adaptation strategy may also be extended to parameter-efficient fine-tuning of visionlanguage models in other domain-specific applications.
Reproduction assets foundThe paper's few-shot crop disease dataset (990 images, 11 crops, 33 classes) is compiled entirely from four public Mendeley Data image repositories, each cited in Table 1 as the data source for specific crop/disease classes. These are the plant image inputs directly used for this paper's phenotyping/analysis. No authorDataset · publicncluding laboratory and field environments (Approximately 45% of them were captured in field environments), to enhance sample representativeness and model robustness.
Table 1
The number of collected diseases or healthy image data for each crop.
Crop
Disease
No. of images
Collection conditions
Data source
Apple
Apple scab
30
Lab
https://data.mendeley.com/datasets/tywbtsjrjv/1
Cedar apple rust
30
Lab
https://data.mendeley.com/datasets/tywbtsjrjv/1
Healthy
30
Lab
https://data.mendeley.com/datasets/tywbtsjrjv/1
Cashew
Healthy
30
Lab
https://data.mendeley.com/datasets/8fr7grr73p/1
Leaf miner
30
Lab
https://data.mendeley.com/datasets/8fr7grr73p/1
Red rust
30
Lab
https://data.mendeley.com/datasets/Open asset ↗lines:37-116Plant phenotyping relevance matchEurope PMC · checked 15 Sept 2026
Existing reviews on AI in tea production are either agriculture-generic or limited to isolated tasks. This review thoroughly compares vision technologies (RGB, hyperspectral, near-infrared, thermal, Light Detection and Ranging (LiDAR), Unmanned Aerial Vehicle (UAV)) and establishes a task-oriented algorithm selection framework for the tea industry. For small-sample or near-linear problems, traditional machine learning (ML) (support vector machine (SVM); partial least squares regression (PLSR)) remains effective. For unstructured field tasks, deep learning achieves superior performance: pest detection accuracy exceeds 97%, tea bud detection reaches 96.8% with RGB images, and hyperspectral imaging predicts nitrogen content with R 2 > 0.90 and tea polyphenols with R 2 up to 0.925. Algorithm choice further differentiates by task granularity: lightweight convolutional neural networks (CNNs) balance speed and accuracy for edge deployment at 16 fps; You Only Look Once (YOLO) series detectors enable real-time localization on mobile platforms at 93.1% accuracy, 24 ms per target. No single algorithm dominates all tea tasks; selection is a trade-off among accuracy, speed, data availability, and computational constraints. These findings outline a structured analysis of the challenges and pathways for transitioning computer vision (CV) from laboratory research toward field-deployable tools.
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.
Summary Understanding how plant populations respond to environmental variation through functional leaf traits remains challenging due to limitations of traditional phenotyping approaches. Hyperspectral reflectance offers a powerful high‐throughput solution, simultaneously capturing leaf biochemistry, water content, and structural properties across hundreds of wavelengths. We present a framework combining hyperspectral data, inverse modeling, and network analysis to investigate population‐level variation in Streptanthus tortuosus . Using a common garden experiment with four populations, we apply supervised methods (partial least square discriminant analysis; ridge regression) to identify which spectral features differ among populations, and an unsupervised spectral network approach to characterize how wavelength correlations are organizationally structured within each population, where we treat coordination architecture itself as a population‐level phenotype that can vary with environment. The framework detects distinct, heritable spectral signatures across populations, population differences in anthocyanins, carotenoids, Chl, water content, and population‐specific network architectures. Thermally variable environments were associated with greater spectral modularity, demonstrating that trait coordination architecture varies with climate of origin. This approach addresses the phenotyping bottleneck in evolutionary ecology, providing a scalable, high‐throughput tool for characterizing genetically based population differences in both individual traits and their coordination, with broad applications for monitoring plant population responses to climate change.
Accurate field detection of candidate tea shoots could support plantation monitoring, yield estimation, fresh-leaf assessment, and future selective-harvesting research. Wuyi rock-tea shoots are small and slender, have weak visual boundaries, and are easily confused with branches, petioles, and complex canopy backgrounds. Here, we developed YOLO11s-CSNG for candidate shoot detection in natural plantation scenes. The model combines a channel-spatial feature enhancement bottleneck, a normalized Wasserstein distance constraint for bounding-box regression, and ghost convolution layers in the detection head. We evaluated the model through detector comparisons, module ablations, and repeated training with five matched random seeds on a natural-scene dataset containing four Wuyi rock-tea cultivars. Across the five matched seeds, the mean mAP@0.5 increased from 67.37 ± 0.91% to 67.97 ± 0.90% on the validation set and from 61.29 ± 0.40% to 61.88 ± 0.66% on the internal test set. Neither paired difference was statistically significant: The 95% confidence intervals included zero, and the exact two-sided paired-permutation p values were 0.375 and 0.250, respectively. The mean mAP@0.5:0.95 did not improve. YOLO11s-CSNG retained a model size and model-only edge-inference time comparable to YOLO11s, providing a compact design for candidate shoot-region detection under the sampled field 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.
Reproduction assets foundThe paper's phenotyping analysis is based on the publicly available PlantVillage plant leaf disease image dataset hosted on Kaggle (54,303 labeled leaf images across 38 classes), which the authors explicitly state was sourced from a publicly available Kaggle dataset. No author-specific code, models, or derived datasetsDataset · publicThe research incorporated PlantVillage dataset(24)
accessible on Kaggle that contains 54,303 plant leaf
images showing both healthy and diseased
conditions spanning across 38 specific categories.Open asset ↗Kagglepdf-raw-page:2 lines:1-105Plant phenotyping relevance matchEurope PMC · checked 15 Sept 2026
Abstract The classification of banana leaf disease has a large impact on agricultural output and relies heavily on timely early detection, with reliability as a fundamental component of effective crop management. The framework proposed in this study comprises a Hamiltonian Full Node Coverage Graph Attention Network (HFNC-GAT) and Fuzzy C-Means super pixel graph learning to explain the classification of banana leaf diseases. The HFNC-GAT allows for the representation of segmented leaf areas as the nodes in a graph. This framework also makes optimal use of an attention learning model to represent the spatial dependence of diseased leaf regions, allowing it to leverage both local and global spatial dependencies. The HFNC-GAT was demonstrated through observed experiments to achieve high performance with 96.11 and 94.19 accuracy, 0.9111 Cohen's Kappa, 0.9111 MCC, 0.9344 F2-score, and 0.9939 ROC-AUC compared to the performance of conventional CNN, GCN, and baseline GAT models.
Reproduction assets foundThe paper's Dataset Availability statement explicitly declares two public Kaggle banana leaf image datasets used for the phenotyping/classification experiments: Banana Leaf Disease Dataset V4 and BananaLSD. No author analysis code or trained model is reported as publicly available.Dataset · publicThe Banana Leaf Disease
Dataset V4 is available at https://www.kaggle.com/datasets/rayhanarlistya/banana-leaf-disease-dataset-v4.Open asset ↗Kaggle · banana-leaf-disease-dataset-v4pdf-page:19 lines:1-55Dataset · publicThe Banana Leaf Spot Diseases (BananaLSD) dataset is available at
https://www.kaggle.com/datasets/shifatearman/bananalsdOpen asset ↗Kaggle · bananalsdpdf-page:19 lines:1-55Code / dataset availability confirmedCrossref · checked 11 Sept 2026
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 11 Sept 2026
Md Nahidur Rahaman · Abdullah Al Mamun · Md. Kamal Hossen · Abdur Rouf · Tumpa Rani Shaha · Jungpil Shin · Mohd Nizam Husen · Abu Saleh Musa Miah
RiceLeafClassificationDisease symptoms / severity
Rice leaf diseases significantly affect crop health and yield potential, creating a need for accurate and timely disease diagnosis. Many existing approaches still rely on single-stream feature extraction architectures, which may limit the ability to simultaneously capture global contextual information and fine-grained disease characteristics. Moreover, limited interpretability and decision-support capability hinder their practical deployment in real-world rice farming. To address these limitations, we employed a framework consists of two parallel feature extraction streams designed to capture different characteristics of disease patterns. The first stream uses a Swin Transformer to learn global contextual information and long-range spatial relationships across the leaf image. The second stream employs ConvNeXt to extract local texture features, including lesion details, spots, and color variations. By combining these complementary representations, the proposed framework effectively integrates global semantic information with local disease-specific features for improved classification performance. The extracted features from the two streams are fused through concatenation followed by an attention-based feature refinement module, enabling adaptive weighting of discriminative features. The refined representation is then used by a fully connected classifier for disease prediction. To enhance model interpretability, Grad-CAM visualization is incorporated to highlight disease-relevant regions and provide visual explanations for the model decisions. Furthermore, an LLM-based advisory module is integrated as a post-diagnosis decision-support component to provide contextualized disease management information and suggestions based on the predicted disease category. The generated suggestions are intended to support, rather than replace, expert agronomic recommendations and should be validated by agricultural professionals before practical application. The proposed framework was evaluated on two rice leaf disease datasets, achieving accuracies of 99.55% and 97.06% on Dataset-1 and Dataset-2, respectively, which are higher than those reported in previous studies. Additionally, five-fold cross-validation on Dataset-1 achieved an average accuracy of 98.91% ± 0.47, demonstrating the stability of the proposed approach. Cross-dataset evaluation using nine common disease classes across both datasets achieved 90.80% accuracy, indicating improved generalization across different data distributions. The proposed framework provides an accurate and explainable approach for rice leaf disease diagnosis in smart agriculture applications.
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.
Nitrogen is an important element present in vital substances such as plant proteins and chlorophyll, the content of which directly reflects the nutrient status of crops, and provides a theoretical basis for crop nutrient diagnosis, growth monitoring and yield potential prediction. Taking the chip-level visible/near-infrared spectral sensor AS7263 as the data acquisition module and the Arduino Uno single-chip microcomputer development board as the control module, a portable crop leaf spectral sensing system was designed in this study. The spectral reflectance and nitrogen content of wheat leaves were obtained through field experiments. Principal Component Analysis (PCA) was used to eliminate abnormal spectral data. Combined with pretreatment algorithms including Multiplicative Scatter Correction (MSC) and Standard Normal Variate (SNV), the prediction models for wheat leaf nitrogen content were established based on Partial Least Squares (PLS), Support Vector Machine (SVR), Random Forest (RF) and a Back Propagation (BP) neural network. The results showed that compared with SNV, the model performance based on the spectral data after MSC pretreatment was better. The test set R 2 values of PLS, SVR, RF and BP models were 0.61, 0.75, 0.83, and 0.89, and the root mean square errors (RMSEs) were 4.62 mg g -1 , 4.38 mg g -1 , 3.39 mg g -1 and 3.27 mg g -1 , respectively. The MSC-BP prediction performance was the best, and the non-destructive and accurate detection of nitrogen in wheat leaves was realized, which verified the feasibility of micro-spectral sensing technology in crop nutrition diagnosis.
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.
O cultivo de maçãs tem grande importância econômica no setor agropecuário brasileiro, especialmente na região Sul do país. No entanto, a produtividade dos pomares é frequentemente comprometida por doenças foliares que, se não tratadas, podem resultar em perdas substanciais. Nesse contexto, os avanços em técnicas de Aprendizado de Máquina têm possibilitado o desenvolvimento de soluções computacionais que auxiliam no diagnóstico agropecuário com maior precisão e agilidade. Este trabalho propõe uma abordagem baseada em Redes Neurais Convolucionais que utiliza aprendizado por transferência para detectar automaticamente sintomas de doenças em folhas de macieira. A metodologia desenvolvida inclui a segmentação e análise de regiões sintomáticas para reduzir o ruído proveniente de áreas saudáveis e direcionar o aprendizado do modelo para sinais relevantes. Um total de 32.382 manchas de sintomas foram extraídas de 1.995 imagens originais, abrangendo cinco classes de distúrbios foliares: glomerela, sarna, danos por herbicidas, deficiência de magnésio e deficiência de potássio. A rede MobileNetV2, treinada por meio de aprendizado por transferência, alcançou um F1-score de 0,926 e 93,8% de acurácia no conjunto de teste reservado. Os resultados indicam um bom desempenho no contexto avaliado, sugerindo o potencial da abordagem como ferramenta de apoio ao diagnóstico da saúde das plantas e à tomada de decisões em campo. Palavras-chave: aprendizado de máquina; visão computacional; doenças em plantas. Application of convolutional neural networks for disease detection in apple tree leaves ABSTRACT: Apple cultivation holds significant economic importance in the Brazilian agricultural sector, especially in the southern region of the country. However, orchard productivity is often compromised by foliar diseases, which, if left untreated, can lead to substantial losses. In this context, advances in Machine Learning techniques have enabled the development of computational solutions that support agricultural diagnostics with greater accuracy and agility. This work proposes an approach based on Convolutional Neural Networks that uses transfer learning to automatically detect disease symptoms in apple leaves. The developed methodology includes segmentation and analysis of symptomatic regions to reduce noise from healthy areas and focus the model’s learning on relevant signals. A total of 32,382 symptom patches were extracted from 1,995 original images, covering five foliar disorder classes: glomerella, scab, herbicide damage, magnesium deficiency, and potassium deficiency. The MobileNetV2, trained via transfer learning, achieved a F1-score of 0.926 and 93.8% accuracy on the held-out test set. The results indicate good performance in the evaluated setting, suggesting the approach’s potential as a tool to support plant-health diagnosis and field decision-making. Keywords: machine learning; computer vision; plant disease.
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 matchOpenAlex · checked 15 Sept 2026
Abstract Crop improvement increasingly depends on resolving how plants move through physiological states during stress rather than where they end up, yet the dominant convention still reduces dense, multi-sensor, longitudinal trait data to a single static value before genetic analysis, discarding information about state, transition and uncertainty. This systematic review asked whether a quantum-inspired, state-based representation of dynamic crop phenotypes could be integrated with established genomic tools to improve trait characterisation and gene discovery, and what the literature reports about the components it would require. Reporting followed the PRISMA 2020 statement and the Synthesis Without Meta-analysis (SWiM) guideline. Scopus, Web of Science Core Collection, PubMed and a Google Scholar grey-literature sweep were searched from January 2017, retrieving 2,413 records, after de-duplication 1,777 titles and abstracts were screened, 279 full texts were assessed, and 82 studies met the eligibility criteria and entered a thematic synthesis. Studies were dual-screened, appraised with an adapted Mixed Methods Appraisal Tool, and their comparable within-study outcomes synthesised by vote counting on direction of effect; meta-analysis was inappropriate because outcomes were not commensurable. Every component of the paradigm probabilistic state representation, temporal trait modelling and trajectory-aware genomic prediction was independently validated, but no included study unified them for crop-stress genetics. All comparable comparisons favoured the temporally richer method, an asymmetry indicating probable reporting bias, and certainty was moderate for representation and modelling and low for realised genetic gain. The phenotyping bottleneck has migrated from measurement to representation.
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.
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.
Automated plant identification based on leaf morphology has gained significant attention in recent years due to its wide range of applications in precision agriculture, biodiversity conservation, environmental monitoring, and botanical informatics. Advances in digital image processing and machine learning have enabled the development of intelligent systems capable of identifying plant species from leaf characteristics with minimal human intervention. Despite these advancements, achieving reliable and accurate classification remains challenging because leaf images are often affected by variations in illumination, complex backgrounds, image noise, differences in orientation and scale, as well as natural leaf deformation. These factors can obscure important morphological features, reduce the effectiveness of feature extraction, and ultimately decrease the accuracy and robustness of automated plant classification systems. Consequently, there is a growing need for intelligent frameworks that can effectively handle these challenges while preserving critical leaf morphology and venation information for reliable plant identification. This study proposes an Intelligent Morphology-Driven Framework that integrates advanced digital image processing and machine learning for robust leaf venation analysis and plant classification. The proposed framework integrates multiple digital image processing and machine learning techniques to enable accurate and automated leaf venation analysis and plant classification. Initially, leaf images undergo preprocessing using grayscale conversion, histogram equalization, Contrast Limited Adaptive Histogram Equalization (CLAHE), Gaussian filtering, Laplacian sharpening, Gabor filtering, and homomorphic filtering to improve image quality and enhance venation and structural details. The enhanced images are then processed through threshold-based segmentation followed by morphological operations, including erosion, dilation, opening, closing, convex hull generation, and skeletonization, to accurately isolate leaf regions while preserving their geometric structure.To characterize leaf morphology, the framework extracts a comprehensive set of features, including geometric descriptors such as area, perimeter, circularity, aspect ratio, solidity, eccentricity, and vein density, together with Hu invariant moments that provide rotation-, translation-, and scale-invariant shape representation. In addition, the framework investigates the influence of image compression by comparing lossless PNG and lossy JPEG formats to evaluate their impact on preserving morphological features and venation details. The extracted feature vectors are subsequently classified using a Random Forest classifier to categorize leaf venation patterns into parallel, reticulate-pinnate, and reticulate-palmate classes.Experimental evaluation demonstrates that the proposed framework achieves an overall classification accuracy of 93.2%, while effectively preserving important morphological characteristics and maintaining computational efficiency. The combination of adaptive image enhancement, morphology-preserving segmentation, comprehensive feature extraction, and robust machine learning classification makes the proposed approach reliable, interpretable, and scalable. Consequently, the framework has significant potential for applications in digital herbarium systems, automated plant identification, biodiversity monitoring, botanical informatics, and precision agriculture.
Abstract Sustainable agriculture has substantial share on improvement of food security and optimization of resources utilization particularly for high value crops like rice leaf. Rice varieties should be properly classified in order to benefit the harvest management, reduced loss after harvest and improved agriculture methods. The traditional classification method usually brings the low precision and the traditional classification method is also subjected to human error, which is difficult to bring about reliable output. This study introduces an optimized multi-class rice leaf disease classification system utilizing “Rice Feature Selection” (RiceFS) and ensemble machine learning approaches. RiceFS is realized based on a feature selection mechanism based on Recursive Feature Elimination. Selected classifiers such as KNN, Random Forest, Gradient Boosting, Ensemble Learning and Optimized SVM are analyzed based on the extracted subset of features and the proposed system is used to classify the seven classes of rice leaf disease. The experimental results show that the Optimized SVM has the best classification results among the different classifiers with accuracy of 92.10%, Precision of 92.20%, balanced Recall and F1 Score, which shows that Optimized SVM is very effective in multi-class rice leaf disease classification. The performance can be improved by feature reduction, generalization capability and computational complexity reduction, which are realized with the help of RiceFS. The proposed framework is designed to provide an intelligent decision support system for timely intervention, loss minimization and sustainable agriculture. Results indicate that these algorithms are applicable for rice leaf disease classification since they are accurate, reliable and scalable.
Reproduction assets foundThe paper's RiceFS phenotyping/classification experiments are built on two public Kaggle rice leaf disease image datasets, explicitly cited with URLs and a data availability statement. No author code or models are deposited.Dataset · publicThe RiceFS framework proposed initially
performs a feature selection, followed by training
several classifiers: K-Nearest Neighbors (KNN),
Random Forest (RF), Gradient Boosting (GB),
Ensemble Learning, and Optimized Support
Vector Machine (Optimized SVM). The data is
published on the Kaggle website. The data is
open-source at:
https://www.kaggle.com/datasets/vbookshelf/rice-leaf-diseases [44].
This data consists of 120 jpgs of disease infected
rice leaves. The photos are divided into 3
categories according to the kind of disease. There
are 40 images in each class.
Classes
• Leaf smut
• Brown spot
• Bacterial leaf blight
The datasets are preprocessed by eliminating
redundant information, normOpen asset ↗Kaggle · vbookshelf/rice-leaf-diseasespdf-raw-page:11 lines:1-103Plant phenotyping relevance matchEurope PMC · checked 14 Sept 2026
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.
Yingchao Wang · Jun Hao · Peng Zhou · Wei Chen · Shan Sun · Na Li · Feng Xue · Zixiang Qin · Hao Wu · Fan Zhao
Aerial / UAVGreenhouseFlowerObject detection
Accurate detection of blooming roses and flower buds is essential for greenhouse phenotyping, cultivation scheduling, harvest planning, and yield management. However, UAV-derived greenhouse imagery presents major challenges because rose targets are often small, densely distributed, partially occluded, and visually similar to complex backgrounds. This study proposes a lightweight rose detection framework that combines multi-source dataset construction with an improved YOLOv12n-based detector, termed BloomRoseNet. A GreenHouse Rose dataset was constructed by integrating self-collected UAV overhead images, screened RoseTracker images, and supplementary multi-view rose images to increase diversity in scale, growth stage, viewpoint, and background complexity. BloomRoseNet introduces task-oriented improvements for fine-grained feature extraction, adaptive feature fusion, and attention-enhanced detection. The supplementary multi-view data improved precision, recall, and mAP@50 from 85.2%, 82.8%, and 89.2% to 86.1%, 85.3%, and 90.5%, respectively. Compared with the baseline YOLOv12n, BloomRoseNet increased precision, recall, mAP@50, and mAP@50:95 by 3.2, 3.3, 3.6, and 1.6 percentage points, respectively, while reducing parameters from 2.55 M to 2.08 M and model size from 5.5 MB to 4.5 MB. The model also maintained real-time inference capability and stronger robustness under blur, occlusion, and illumination disturbances. The proposed framework provides an effective and practical solution for UAV-based greenhouse rose monitoring and supports precision cultivation management.
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:1-53Code / dataset availability confirmedEurope PMC · checked 5 Sept 2026
El Karch H, Natij Y, Benaly M, El Gouri R, Mezouari A.
Field / plotLeafClassificationStress / disease detectionDisease symptoms / severity
Automated plant disease recognition from field photographs remains challenging: models trained on laboratory datasets collapse on in-the-wild images, and the current state of the art on PlantWild-the largest open in-the-wild benchmark (18,542 images, 89 classes)-relies on text prototypes derived from a language model to reach 76.18% top 1 accuracy. We ask whether text supervision is truly necessary or whether the bottleneck is the diversity of the visual representation. Our central methodological finding is that backbone selection and fusion matter far more than classifier-head engineering : across 118 experiments, a simple fixed-weight linear-prototype combination on top of three complementary frozen backbones yields larger and more reproducible gains than any head-level adaptive routing mechanism we test. Specifically, through a systematic study of six frozen vision backbones (three CLIP, two DINOv3, and one DINOv2), three classifier heads, and four fusion configurations, completed in a single day on one consumer GPU, we establish three findings. (i) A single self-supervised backbone (DINOv2 ViT-L/14) already surpasses text-augmented MVPDR (77.56% vs. 76.18%). (ii) Concatenating three complementary backbones (DINOv2 + DINOv3 + CLIP) reaches 80.23% ± 0.41% (five seeds), exceeding the published MVPDR accuracy by +4.05 points and our own reproduction of MVPDR under an identical evaluation protocol by +7.96 points, without any language supervision. (The difference between the two deltas reflects evaluation-protocol differences-our split, model-selection criterion, and training schedule-rather than any discrepancy in the reported numbers; see Section 4.6.6 for a full reconciliation.) (iii) Every form of learned routing we test-per-class gating, backbone gating, sample-wise gating-is inessential; the gain is entirely attributable to backbone diversity and a simple linear-prototype scoring combination. On the smaller PlantDoc benchmark, the same principle transfers but with substantially higher seed variance: the best configuration reaches 80.09% at a favourable seed but 76.97% ± 1.48% over five seeds-a suggestive rather than robust gain. Beyond the accuracy headline, we provide a pathology-aware per-class analysis showing that DINOv2/v3 dominate on fine-texture lesion classes (rusts, mildews, and leaf spots) whilst CLIP's narrow advantage concentrates on organ/species-level identification (rice leaf and potato late blight). All primary claims are validated over five seeds, and all code, feature caches, and result files are released for full reproducibility.
Reproduction assets foundThe paper states that its code, cached features, and result files will be released, but no authors' public URL or repository is provided in the supplied blocks (future-tense availability language only). The DINOv3 GitHub link and OpenReview link are third-party backbone resources, not paper-specific assets.Code · publicAll primary claims are validated over five seeds, and all code, feature caches, and result files are released for full reproducibility.Open asset ↗lines:332-335Code / dataset availability confirmedCrossref · checked 14 Sept 2026
Published11 Aug 2026Engineering, Technology & Applied Science ResearchCited by 0 · OpenAlex ↗
In recent years, areca nut plants have been vulnerable to different diseases that appear as distinct colors on leaves, caused by bacteria or fungi. These symptoms disrupt photosynthesis and reduce yield, affecting productivity and crop health. Therefore, accurate plant disease classification is essential for detecting distinct disease shapes and sizes. Existing Deep Learning (DL) models have several limitations that prevent them from distinguishing between various plant diseases due to similar characteristics. To overcome this limitation, a Dynamic elastic boundary strategy Sailfish Optimization Algorithm and Convolution Neural Network (DSFO-CNN) method is proposed to identify and accurately classify arecanut plant diseases. The Visual Geometry Graph-19 (VGG-19) model extracts features that have significant information about disease in arecanut plants. The proposed arecanut plant disease classification model employed feature selection and drop cyclic learning rate, which adjusts the CNN learning rate to efficiently learn the subtle information about various leaf and nut diseases to enhance classification. The experimental results of the DSFO-CNN demonstrate superior performance compared to existing approaches.
Reproduction assets foundThe paper's phenotyping inputs are two public image datasets: the collected Arecanut dataset (Kaggle) and the PlantVillage dataset (Kaggle), both explicitly cited and declared openly available. No author analysis code or trained model is released.Dataset · publicDATA AVAILABILITY
The data used in this study are openly available at [19] and
[20].Open asset ↗pdf-page:7 lines:1-63Plant phenotyping relevance matchCrossref · Europe PMC · checked 15 Sept 2026
Accurate crop disease diagnosis is critical for agricultural productivity and food security, yet existing deep learning systems often struggle to generalize across visually similar diseases and varying environmental conditions. Recent Vision-Language Models (VLMs) have demonstrated promising zero-shot reasoning capabilities; however, most agricultural diagnostic systems still rely on isolated single-model predictions without collaborative reasoning or consensus mechanisms. In this work, we propose VIDA+PANDA, a multi-agent Vision-Language framework for zero-shot crop disease diagnosis. The framework consists of two stages: VIDA, where multiple VLM agents independently analyze crop leaf images to establish baseline performance, and the Peer-Anchored Named Deliberation Architecture (PANDA), which introduces a structured multi-round debate among a selected group of high-performing and architecturally diverse agents. During deliberation, agents exchange reasoning, critique peer predictions, and revise decisions through evidence-grounded discussion, while an anti-sycophancy mechanism discourages unsupported consensus shifts. Final predictions are generated through performance-weighted consensus voting. Experiments are conducted on the CDDM benchmark using seven heterogeneous VLMs from four independent providers, including two open-source models, under a fully zero-shot setting. A non-participant GPT-5 model serves as an independent judge to assess the final diagnostic predictions. Beyond conventional accuracy, the framework introduces three semantic measures: Semantic Label Similarity (SLS), which measures how semantically close a predicted crop-disease pair is to the ground truth and captures partial correctness overlooked by exact-match evaluation; Reasoning Specificity (RS), which measures how concretely an agent’s explanation references visual evidence such as lesion color, shape, texture, or margins; and Inter-Agent Reasoning Convergence (IRC), which measures the extent to which agents rely on similar visual evidence, capturing epistemic alignment independently of label correctness. Experimental results show that collaborative multiagent deliberation improves individual diagnostic performance and semantic alignment, with the largest gains observed among weaker participating agents. The findings also reveal important relationships between predictive accuracy, persuasive influence, and consensus formation in VLM-based agricultural diagnosis.
Joice A, Tazeen H, Tufaique T, Juliet Gupta M, Igathinathane C, Rai N, Whippo CW, Archer DW.
MaizeField / plotClassificationGrowth / development / phenology
Integrating precision agriculture (PA, a data-driven agricultural management system) with deep learning (DL) models can effectively support various activities, including yield prediction, crop health monitoring, field task automation, and decision-making. Taking advantage of such data-driven methodologies typically requires desktops, high-performance computing systems, and cloud clusters for data analysis, but their portability limits in-field applications. However, a single board computer, such as Raspberry Pi, offers a compact, lightweight, cost-efficient, easy-to-use, and feature-rich portable computing device which is ideal for in-field decision-making in PA applications. One such in-field application is crop growth stage classification for better crop management. Therefore, in this study, eight corn growth stages were classified using PhenoCam (near-surface [proximal] remote sensing network camera)imagery collected from ten PhenoCam sites. Four lightweight DL models were developed, ELiteCrop0, ELiteCrop1, ELiteCrop4, and MobNetCropV2, and evaluated across five image vertical clipping levels(0 %–40 %) using a supercomputer. The optimized model was subsequently deployed on a Raspberry Pi5 for edge inference. Model training accounted for the majority of the total CPU time, exceeding 97 %, while the testing times ranged from 0.01 min to 0.12 min, enabling real-time applications. Among the models, ELiteCrop0 achieved the most balanced performance with a confusion-matrix diagonal ratio(CMDR) of 0.93, followed by ELiteCrop1 (CMDR = 0.92). Overall, model performance decreased with increasing vertical clipping; therefore, a moderate image clipping (0 %–10 %) was recommended for improved computational efficiency. Analysis with a supercomputer produced an intrasite (same train sites)accuracy of 0.90–0.93 (Raspberry Pi: 0.78–0.81) and an intersite (new test sites) accuracy of 0.48–0.50(Raspberry Pi: 0.41–0.43), indicating challenges with model generalization. Raspberry Pi successfully processed ≈ 1000 images/min under safe operating conditions (68◦C). Future work should focus on extending the multi-site dataset to improve cross-site performance. Hence, this study presents a scalable and cost-effective solution for real-time corn growth stage monitoring in PA.
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:419-444Plant phenotyping relevance matchEurope PMC · checked 5 Sept 2026
Germination percentage is an endpoint measure and therefore does not describe when an individual seed begins visible growth or how rapidly its radicle and plumule expand. We developed a time-resolved phenotyping workflow to quantify rice seed germination continuously in shallow-water culture. A single industrial camera moved along a 1 m rail and imaged three culture boxes at 1 h intervals for up to 80 h. The archive comprised 1,062 full-frame images and 6,372 seed-level repeated observations under the six-seed field-of-view configuration. A physical grid maintained seed identity through time and enabled individual regions of interest to be extracted. Whole-seed foregrounds were obtained with a pretrained U 2 -Net, and a masked RGB intensity rule separated newly emerging tissue from the darker hull. For each tracked seed, projected emerging-tissue area and interval growth rate were calculated. Three representative normally germinating seeds first showed measurable tissue at 48 h, yet subsequently followed distinct trajectories: final projected areas ranged from 2,605 to 4,700 pixels and peak interval growth rates ranged from 106.88 to 287.92 pixels h −1 . B-1 accumulated 63.71% of its final visible area during 72–80 h, whereas B-3 accumulated 73.51% during 60–72 h. Thus, seeds with the same observed emergence interval can differ substantially in the timing and magnitude of post-emergence expansion. The workflow converts repeated images into biologically interpretable temporal phenotypes and provides a basis for nondestructive studies of rice seed vigor and germination heterogeneity.
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.
Background: Plant diseases significantly threaten global food security, reducing potential harvests and sometimes causing total crop failure. Traditional detection methods, which rely on manual inspection and laboratory testing, are time-consuming, costly and prone to human error. Methods: To address these challenges, this study applies transfer learning techniques using deep convolutional neural networks for accurate and efficient plant disease detection. A comparative analysis of nine pretrained models, VGG16, VGG19, ResNet50, ResNet101V2, MobileNetV2, InceptionV3, DenseNet121, InceptionResNetV2 and Xception was conducted on the PlantVillage dataset, focusing on apple, potato and peach leaf images. Result: Results show that DenseNet121 and ResNet101V2 achieved the highest accuracy, particularly for potato leaves with 98.5%, while MobileNetV2 also performed well with up to 99% accuracy for apple and peach leaves. The study demonstrates that transfer learning effectively enhances plant disease classification, enabling faster, more reliable and resource efficient detection for precision 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.
Field / plotLiDAR / point cloudStem / branchYield / biomass estimationBiomass / plant weight
Integrating ground-based and aerial remote sensing for individual tree-level stem volume modeling remains underexplored in Mediterranean mixed forests, despite the growing need for cost-effective, automated forest inventory approaches. This study evaluated the combined use of Handheld Laser Scanning (HLS) and Unmanned Aerial Vehicle (UAV)-based Structure from Motion (SfM) photogrammetry for individual-tree stem volume modeling in a mixed stand in Castilla y León, Spain, dominated by Pinus halepensis, Pinus pinea, Quercus faginea, and Cupressus sempervirens. Two open-source HLS processing tools; the Forest Structural Complexity Tool (FSCT) and 3D Forest Inventory (3DFin), were compared for individual tree attribute extraction, with FSCT outperforming 3DFin across all species. Reference stem volumes were derived by applying species-specific Spanish National Forest Inventory (SNFI) allometric equations to FSCT-extracted diameter and height values. Random Forest models were then built using UAV-SfM crown metrics as predictors, testing two image overlap configurations: 80 × 80 F (80% front and side overlap) and 80 × 60 CF (80% front, 60% side, cross-flight). The 80 × 80 F configuration produced the best-performing model (R2 = 0.730), with 80 × 60 CF achieving comparable accuracy (R2 = 0.688), results confirmed by spatially independent leave-one-plot-out cross-validation (LOPO-CV R2 = 0.627 and 0.613, respectively). These results show that combining HLS and UAV-SfM through a predominantly open-source workflow offers a viable, reproducible approach to stem volume modeling in structurally complex Mediterranean mixed forests.
The tomato plant is considered one of the most important crops in the world, yet it is vulnerable to various diseases that affect crop quality and agricultural productivity. These challenges have driven the need for an efficient and intelligent plant disease detection system. With the development of computer vision and artificial intelligence, this proposed methodology based on deep learning for tomato leaf diseases has been presented. Two public datasets: Taiwan DS with nine classes and Tomato Leaf Diseases Detection Computer Vision Dataset (TLDDCV DS) with seven classes have been used to test this system. This system begins with plant image processing, which includes gamma correction and bilateral filtering, to enhance image quality and clarity while preserving key disease features. Then, a genetic metaheuristic algorithm was used to automatically select the most significant hyperparameters, further optimizing both processing time and accuracy. After that, the tomato leaf disease detection applies the You Only Look Once version 11 Nano (YOLOv11n) model. The YOLOv11n backbone is edited through a Data-efficient Image Transformer (DeiT) to improve the system's capacity for learning global contextual information and long-range dependencies. Experimental results demonstrate that the proposed system outperforms existing methods. It achieved an average mAP@50 of 97.8%, mAP@50-95 of 93.4%, precision of 97.3%, recall of 93.8%, and F1-score of 95.5% on the Taiwan dataset. Additionally, it achieved an average mAP@50 of 87%, mAP@50-95 of 48%, precision of 83.9%, recall of 70.3%, and F1-score of 76.4% on the TLDDCV dataset. These results demonstrate the generalizability and effectiveness of the proposed system in real-world agricultural situations.
Reproduction assets foundThe paper uses two public Roboflow tomato leaf disease image datasets and states its source code is publicly available on Zenodo, all with explicit availability statements and URLs.Dataset · publicThe first dataset is the Taiwan dataset, which can be found at the following link: (https://universe.roboflow.com/bryan-b56jm/tomato-leaf-disease-ssoha).Open asset ↗tomato-leaf-disease-ssohalines:317-328Dataset · publicThe second dataset is the TLDDCV dataset, which can be found at the following link: (https://universe.roboflow.com/sylhet-agricultural-university/tomato-leaf-diseases-detect)Open asset ↗tomato-leaf-diseases-detectlines:317-328Code · publicThe source code of the proposed framework, including the implementation of the proposed methodology and experimental setup, is publicly available in the Zenodo repository: https://doi.org/10.5281/zenodo.20777853 .Open asset ↗Zenodo · 10.5281/zenodo.20777853lines:317-328Plant phenotyping relevance matchCrossref · checked 5 Sept 2026
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.
Brassica juncea (Mustard) is one of the most economic seed vegetable crops of the world, playing a major role in the production of world edible oil and the agricultural economy. The third-largest producer is India, which had an area under cultivation of about 8.6 million hectares of Mustard in 2021 22, and annual revenue of over USD 5 billion in 2021 22. Nevertheless, the presence of diseases like Alternaria Leaf Spot, White Rust, Powdery Mildew, and Septoria Leaf Spot threatens yield and quality by up to an estimated 2070% loss every year, based on the severity of the disease, and thus economic losses are estimated at over USD 1.5 billion per year in India alone. Traditional diagnostic systems are based on a manual examination of an agronomist trained to look at the sample and make a judgment, which is time-consuming, subjective, and subject to human error. Current deep learning methods of automated disease detection, promising as they are, are prone to inaccuracies on complex disease patterns, poor uncertainty estimation that is essential in real-world implementation, and poor generalizability to different field conditions. In response to these drawbacks, Swin-BNN-RF, a hybrid framework that combines Swin Transformer as a hierarchical attention-based feature extractor, Bayesian Neural Network (BNN) with symmetrized posterior as a probabilistic learner, and a Random Forest (RF) as an ensemble classifier, is proposed in this study. The Swin Transformer also harnesses local and global spatial biases with its shifted window self-attention network, and it is able to extract better features on leaf images of high-resolution. The uncertainty estimates of the BNN component are trusted, and unambiguous predictions are highlighted to get the opinion of the human expert. Random Forest classifier uses the bagging and boosting ensemble methods to improve stability and the robustness of the classification. A large dataset was experimented with; it consisted of more than 10,000 samples per category of disease in four diseases. In the case of binary classification, the proposed model was 98.32% accurate, 98.52% precise, 98.70% recall, and 98.36% F1. On multi-class classification, it obtained 97.50, 97.82, 98.51, and 97.46 accuracy, precision, recall, and F1 score, respectively, which showed consistent performance in comparison with state-of-the-art models such as EfficientNet, MobileNet, and Residual Networks. The contribution of each component is verified by the Ablation studies and statistical analysis of significance (p
Accurate detection of small-target diseases on apple leaves is of great importance for optimizing orchard management and facilitating precision agriculture. Small-target disease detection remains challenging due to insufficient features and complex backgrounds. This work proposes an effective detector for apple leaf small-target diseases called ALSDet. The global context module is integrated into Stage2 to Stage4 of the ResNet-50 backbone, yielding a refinement of the feature-extraction architecture. In the bottleneck blocks of the Stage2 and Stage3, dilated convolution is used in place of the normal 3×3 convolution to enlarge the receptive field for small targets and strengthen feature extraction. During the model training phase, a multi-scale training strategy combined with the online hard example mining method is adopted to focus on learning hard samples and enhance adaptability for various scale targets. According to the experimental results, ALSDet obtains a mean average precision (mAP) of 65.6% and an average recall (AR) of 71.2% on the dataset. The proposed model achieves the highest levels in both mAP and AR when compared to the popular object detection models, such as Cascade R-CNN, Faster R-CNN, GFL, Grid R-CNN, Libra R-CNN, FCOS, VFNet, RetinaNet, SSD, YOLOv7, and YOLOv8. For small-target diseases like rust and frog eye leaf spot, the average precision surpasses 87% with an intersection over union (IoU) threshold of 0.5. These results confirm that ALSDet achieves stable performance against existing methods, demonstrating its potential as a practical tool for intelligent orchard disease management.
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.
Tolerance against winter freeze is the main focus of variety development in Louisiana, which represents the northernmost sugarcane-growing region worldwide. Antifreeze metabolites, xylem structure, and fiber content represent interrelated physicochemical properties contributing to freeze tolerance. This study first classified the cold tolerance of sugarcane cultivars using metabolites in juice as predictor variables. The best-fit model (XGBoost discriminant analysis) estimated the higher cold tolerance of the final on-station clone progeny to the stress tolerance-inducing wild germplasm line. Stalks of the tolerant sugarcane genotype contained higher fiber for mechanical support against cellular injury, compared to susceptible varieties. Fluorescence microscopy visualized phospholipids responsible for maintaining membrane fluidity during frost in lignin surrounding the vascular bundle. Thermal imaging is proposed for real-time monitoring of spatiotemporal temperature changes, as stalk injury is initiated by ice formation at sub-freeze temperatures during winter freeze. As additional datasets for independent prediction become available, developed methods could be used to explore the biomarkers for stress resistance in simpler multivariate discriminant analysis and the distribution of specific biomarkers in cellular components by microscopic imaging, and to trace stalk injury hot spots as a function of time and relationships with resistance markers.
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.
Early and accurate plant disease detection is essential for reducing crop losses and supporting sustainable agricultural management. Although deep learning-based approaches have achieved strong performance in plant disease analysis, many existing models require substantial computational resources, which limits their use in resource-constrained agricultural environments. This study proposes ShuffleNetV2-hSimKD, a lightweight integration framework for plant disease detection. It adopts ShuffleNetV2 as the backbone and incorporates the parameter-free SimAM attention mechanism to enhance disease-related feature representation without introducing additional learnable parameters. In addition, the standard ReLU activation function is replaced with h-swish to improve nonlinear feature extraction and preserve informative feature responses. A hybrid knowledge distillation strategy is further employed to transfer both output-level and feature-level knowledge from a high-capacity teacher model to the lightweight student network during training. Unlike previous studies that apply these techniques in isolation, ShuffleNetV2-hSimKD synergistically integrates parameter-free SimAM, h-swish optimization, and hybrid KD to overcome the representation limitations of lightweight backbones in subtle disease symptom detection. The proposed framework was evaluated on a balanced subset of the PlantVillage dataset, in which leaf images were categorized as healthy or diseased. ShuffleNetV2-hSimKD achieved an accuracy of 90.41% with only 1.4M parameters and 151M FLOPs. Compared with representative lightweight Convolutional Neural Networks (CNNs), the proposed model achieved improved accuracy and recall while maintaining low computational complexity. These results demonstrate that ShuffleNetV2-hSimKD provides an effective balance between detection performance and computational efficiency, highlighting its potential as a lightweight candidate for plant disease detection in resource-constrained agricultural scenarios.
Introduction The precise detection of microscopic tea leaf diseases is a prerequisite for sustainable precision agriculture. While recent deep learning advancements often favor architectural complexity, this "complexity bias" frequently introduces computational redundancy-a "complexity tax"-that destabilizes gradient flow and fails to resolve critical resolution bottlenecks for micro-lesion identification. Methods We propose Opti-YOLOv11n, a minimalist optimization paradigm prioritizing physical input fidelity and stabilized gradient dynamics. Utilizing a dataset of six pathological categories, our framework employs high-resolution scaling (832×832) combined with a momentum-based SGD optimizer and a cosine annealing schedule to reconstruct essential spatial textures. Results Opti-YOLOv11n achieved a peak Precision of 98.87% and a Recall of 95.97%, while reducing the parameter count to 2.35 M-a 9.2% decrease relative to the baseline-and maintaining a real-time inference speed of 104.7 FPS on edge-simulated hardware. Discussion Statistical verification via 5-fold cross-validation confirms superior generalization stability. These results substantiate that strategic structural pruning and physical input scaling provide a more robust technical benchmark for autonomous plant protection than the adoption of excessive architectural depth.
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.
Abstract Bacterial fruit blotch (BFB) caused by Acidovorax citrulli , is a destructive seed-transmitted disease that seriously threatens global cucurbit production. To address the need for detecting A. citrulli -infested seeds, this study developed a colony identification model and a seed infestation detection model based on multispectral imaging. The combined nMahalanobis and nCDA colony identification models achieved a high recall of 0.999 and a low false-positive rate of 0.149 when tested on samples. For infested melon seed detection, we evaluated and compared the classification performance of seven machine learning models. The results showed that LDA, logistic regression, and MLP exhibited stable performance on artificially infested seed samples. Furthermore, multi-cultivar modeling improved model generalizability and demonstrated the feasibility of using multispectral imaging to identify naturally infested seeds. When a qPCR Ct threshold of 37 was used to define seed infestation status, the logistic regression model achieved a validation accuracy of 0.82. Overall, these findings demonstrate the potential of multispectral imaging for colony identification and seed infestation detection, providing a new technical approach and a scientific basis for seed health testing of bacterial fruit blotch in cucurbit crops.
Leaf chlorophyll content (LCC) is a key indicator for assessing the photosynthetic capacity and nutritional status of winter wheat. Among traditional LCC estimation methods, empirical models lack a physical basis and have poor generalisability, while physical models are widely applicable but suffer from ill-posed inversion problems. Hybrid inversion methods, which integrate radiation transfer models such as PROSAIL with machine learning, offer both the interpretability of physical models and the efficiency of machine learning; however, they are still affected by the domain shift between simulated and measured data, which limits their generalisation performance. Transfer Component Analysis (TCA), a domain adaption method, can effectively alleviate this problem. In this study, hyperspectral and LCC data were collected in the field, and simulated data were generated using the PROSAIL model; a sensitivity analysis was conducted to identify LCC-sensitive bands. A genetic algorithm was applied to the measured data for band selection and, together with the results of the sensitivity analysis, yielded an optimal set of 30 characteristic bands for subsequent modelling. Three datasets were constructed: measured data only, a direct mixture of measured and simulated data, and a TCA-fused mixture of measured and simulated data. Four models-gradient boosting regression (GBR), random forest (RF), support vector regression (SVR) and deep neural network (DNN)-were developed for each dataset. The results show that: (1) the LCC-sensitive bands are concentrated in the 450-660 nm and 680-720 nm ranges; (2) the model built on the TCA-fused data (R² = 0.722, RMSE = 6.792) outperformed those built on the measured-only data (R² = 0.682, RMSE = 7.259) and the directly mixed data (R² = 0.616, RMSE = 7.976); (3) for the TCA-fused data, the four models differed considerably in accuracy, with SVR performing best (R² = 0.723, RMSE = 5.363), followed by RF (R² = 0.630, RMSE = 6.201) and GBR (R² = 0.575, RMSE = 6.650), whereas the DNN performed worst (R² = 0.388, RMSE = 7.947), probably owing to the limited sample size.
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.
Precise and reliable diagnosis of leaf diseases in tomato is essential for enhancing crop cultivation and minimizing agricultural losses. While deep learning models have performed well on benchmark datasets, the majority of present techniques rely on flat multi-class classification, which predicts all disease categories simultaneously. Such formulations promotes inter-class confusion, particularly when biologically different diseases with similar visual symptoms are learned within a same model. To overcome this constraint, we propose a biologically structured hierarchical deep learning framework in this study. Instead of directly classifying 10 disease classes, the proposed method first classifies leaf images into meaningful biological groups such as bacterial, fungal, pest-associated and healthy using a vision transformer (ViT) model. Then, specialized convolutional neural network (CNN) experts perform fine-grained classification within each category. The proposed hierarchical model shows an overall accuracy of 97.8%, when validated on PlantVillage tomato dataset. A flat ViT model trained with class-weighted loss obtained 96.3% accuracy, whereas a flat CNN model reached 99.3% under clean conditions but decreased sharply to 41% under Gaussian perturbation ( σ = 0.05). On the other hand, the hierarchical model performed steadily under noise with 97.4% accuracy at the same perturbation level. These results indicate that adding biological structure to model design reduces confusion, helps prevent imbalance effects and increases robustness, providing a more trustworthy and interpretable solution for real-world agricultural disease diagnosis.
Reproduction assets foundThe paper's phenotyping/classification measurements are based on the public PlantVillage tomato leaf image dataset, which the authors explicitly state was analyzed and provide a public Kaggle URL. No author analysis code, trained models, or paper-specific supplementary assets are described with availability language.Dataset · publicIsabel Luna-Maldonado , Autonomous University of Nuevo León, Mexico
Reviewed by: Noredine Hajraoui , Moulay Ismail University, Morocco
Tri Handhika , Universitas Gunadarma Pusat Studi Komputasi Matematika, Indonesia
Data availability statement
Publicly available datasets were analyzed in this study. This data can be found here: https://www.kaggle.com/datasets/charuchaudhry/plantvillage-tomato-leaf-dataset .
Author contributions
HG: Conceptualization, Formal analysis, Methodology, Writing – original draft, Writing – review & editing. SR: Supervision, Validation, Writing – review & editing. BL: Supervision, Validation, Writing – review & editing.
Conflict of interest
The author(s) declared thaOpen asset ↗Kaggle · charuchaudhry/plantvillage-tomato-leaf-datasetlines:588-616Plant phenotyping relevance matchCrossref · checked 5 Sept 2026
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.
Accurate, field-deployable diagnostic tools are needed to close the diagnostic gap that limits fungicide targeting among smallholder tomato farmers in Northern Nigeria. This study developed and validated a lightweight convolutional neural network (CNN) for automated diagnosis of five major tomato fungal diseases plus three additional common conditions, trained on field-collected leaf images from Kano and Kaduna States. A MobileNetV2 architecture pre-trained on ImageNet was fine-tuned via transfer learning on more than 10,000 images across ten disease and health classes, using farm-level dataset splitting to prevent data leakage and five-fold cross-validation for model selection.
Abstract Background Leaves maintain hydraulic homeostasis during photosynthesis through the coordinated action of stomata, which regulate gas exchange and transpiration, and veins, which supply water to the leaf lamina. While functional links between stomatal and vascular traits are known in dicots, their potential genetic coordination in C4 crops remains poorly understood. We investigated the genetic architecture of these traits in maize using a Multi-parent Advanced Generation Inter-Cross (MAGIC) population and a low-cost, high-throughput phenotyping platform integrating leaf clearing, digital microscopy, artificial intelligence, and image analysis Results We phenotyped 285 recombinant inbred lines and the MAGIC founder lines, generating 8,072 images from 2,026 leaf samples taken from seedlings grown in controlled conditions. A YOLOv8-based model automatically detected stomata, while a custom and efficient image-processing pipeline quantified vein traits and stomatal spatial distribution patterns along cell bundles. This enabled simultaneous characterization of stomatal density, size, and distribution together with vein density, thickness, and bundle-associated spatial patterning. Substantial phenotypic variation was observed among genotypes, with strong correlations between abaxial and adaxial traits but no significant correlations between stomatal and vein traits. QTL mapping identified 37 genomic regions associated with stomatal and vein traits, including loci containing known developmental regulators such as stomatal density and distribution1 and stomagen1 , as well as novel loci controlling stomatal spatial patterns, divergence between leaf surfaces and veins traits. Conclusions These results support independent genetic control of stomata and veins and decoupled contribution to water-use efficiency, providing a novel genetic framework to independently optimize leaf hydraulic capacity and gas exchange in target environments.
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.
Introduction This study evaluates the applicability of digital morphometric analysis to images of archaeological carbonized olive endocarps as a proof-of-concept initial approach for archaeobotanical investigations. Conventional morphometric analyses of olive endocarps largely rely on manual measurements, limiting reproducibility and quantitative comparison. Methods Ten archaeological endocarps were selected from previously published archaeological assemblages based on the integrity of their outlines, apex-base morphology and overall preservation quality. Quantitative descriptors describing endocarp size, symmetry, curvature and contour geometry were extracted using the OliveID software and compared with a modern morphometric reference database comprising Greek and international olive cultivars. Results Reliable contour extraction and quantitative descriptor computation were successfully achieved for all archaeological specimens despite carbonization. Preliminary comparison of representative morphometric descriptors showed that the archaeological specimens were positioned within the morphometric variation observed among the modern reference collection. Hierarchical clustering consistently associated the archaeological endocarps with the modern Throumbolia morphotype, while distinguishing them from elongated, globular and mucro-bearing cultivars. Discussion These findings demonstrate the feasibility of applying digital image-based morphometric analysis to sufficiently preserved archaeological carbonized olive endocarps and indicate a similar morphometric affinity between the analyzed archaeological material and the modern Throumbolia cultivar. This proof-of-concept study highlights the potential of digital morphometric approaches for quantitative archaeobotanical investigations of archaeological olive remains, while emphasizing the need for larger archaeological datasets and standardized image acquisition to further validate the observed morphometric similarity.
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 quantitative measurements of the archaeological specimens are presented in Supplementary Table 2 , whereas the corresponding mean values and standard errors for the modern cultivars are provided in Supplementary Table 3 .Open asset ↗lines:311-320Code / dataset availability confirmedEurope PMC · checked 5 Sept 2026
The simultaneous diagnosis of diseases and evaluation of age quality grades in tea leaves are critical for precision agriculture and the economic valuation of tea products. Although deep learning has shown promise in agricultural vision tasks, current multi-task models often suffer from performance degradation due to feature conflicts: tea leaf disease recognition relies heavily on macro-structural lesions, whereas tea leaf-age quality grading depends on micro-textural features such as trichome density and color uniformity. To address this discrepancy, we propose a novel dual branch fusion network. Our architecture fundamentally decouples the feature extraction process by utilizing a dual branch mechanism. The first branch employs global average pooling to capture first-order spatial statistics; it can retain the global structural layout necessary for macro-lesion detection. The second branch introduces a dimensionality-reduced self-bilinear pooling module to compute second-order covariance matrices; it can effectively capture the fine-grained textural patterns essential for micro-grade classification. These decoupled features are subsequently fused and optimized through a weighted multi-task loss function. Experimental results on a comprehensive tea leaf dataset demonstrate that the proposed dual fusion framework significantly outperforms baseline models. The proposed network can rescue the disease classification accuracy drop observed in standard bilinear models while maintaining exceptional grading performance. Furthermore, the proposed network maintains a compact parameter footprint and low computational complexity. This balance renders it suitable for deployment on agricultural Internet of Things edge devices where inference speed is critical.
Reproduction assets foundThe paper's Data Availability Statement publicly releases the two tea leaf image datasets used for its phenotyping tasks (disease recognition and leaf-age quality grading) via Mendeley Data. The authors' analysis code and trained models are only promised 'upon acceptance' with no public URL, so they do not qualify.Dataset · publicThe tea leaf disease recognition dataset analyzed in this study is available from https://data.mendeley.com/datasets/744vznw5k2/3 (accessed on 11 February 2026)Open asset ↗744vznw5k2/3lines:514-565Dataset · publicthe tea leaf grading dataset is available from https://data.mendeley.com/datasets/7t964jmmy3/1 (accessed on 11 February 2026)Open asset ↗7t964jmmy3/1lines:514-565Code / dataset availability confirmedEurope PMC · checked 5 Sept 2026
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 matchEurope PMC · checked 15 Sept 2026
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.
Reliable traits are needed for identification of tea ( Camellia sinensis ) cultivars, yet the stability of leaf morphology and color across leaf positions remains unclear. This study evaluated inter-cultivar variation and positional stability in leaf morphological, RGB color, and SPAD traits in six predominant cultivars. One-year-old shoots were sampled in a completely randomized design, and five fully expanded leaves below the apical bud were analyzed. SPAD values were measured with a chlorophyll meter, and scanned images were used to extract contour and RGB traits. Data were analyzed using ANOVA, correlation analysis, PCA, and discriminant analysis. Leaf morphology differed among cultivars and leaf positions, with significant cultivar-by-position interactions; however, the width-to-length ratio differed among cultivars but remained stable across positions in these cultivars. SPAD values increased with leaf position and were strongly associated with RGB components, being negatively correlated with R and G and positively correlated with B. Morphological traits explained 52.988% of total variance in PCA and yielded 64.6% overall classification accuracy, with LaoHan showing the highest accuracy (83.3%). Misclassification was concentrated among genetically similar cultivars. These findings suggest that stable leaf shape proportions and SPAD-RGB relationships provide useful descriptors, whereas genetic relatedness limits morphology-based cultivar identification under the present conditions.
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 following supporting information can be downloaded at https://www.mdpi.com/article/10.3390/biology15151283/s1 , Table S1: Original data of leaf morphological traits, RGB values, and SPAD values from six tea cultivars in this study.Open asset ↗lines:368-409Code / dataset availability confirmedarXiv · OpenAlex · checked 15 Sept 2026
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 matchEurope PMC · bioRxiv · checked 15 Sept 2026
1 Summary Embolism, the formation of air bubbles in the plant water transport system, is a mechanistic driver of plant death. The Optical Vulnerability Technique (OVT) is an imaging method for non-invasive quantification of embolism (including P50, a common metric for drought vulnerability), which can also provide detailed spatial and temporal information. Its major cost lies in the post-processing of thousands of images. Here we designed, tested, trained, and make publicly available a neural network model to automate post-processing of OVT images. Using a dataset of 65 leaves from Senecio pterophorous , we compared our model predictions to results obtained via traditional post-processing by an expert. Our model resolved P50 to within 0.027 MPa of the expert-processed data with training taking 30 minutes to 2.5 hours and model-runtime in the order of seconds to minutes, demonstrating its promise for increasing the efficiency and throughput of P50 calculation. The model’s performance in replicating the pixels that constitute embolism events was lower (mean event-frame IoU of 0.38). We invite the community to utilise our model but emphasise that it does not replace the expert-processing pipeline and that care must be taken when considering applying this and similar approaches to OVT data.
Accurate, non-destructive assessment of watermelon ripeness remains a significant challenge in horticultural production, particularly under field conditions where traditional visual and tactile evaluation methods are subjective and often inconsistent. Although mechanical, acoustic, and spectroscopic techniques have demonstrated promising performance, their reliance on controlled laboratory environments limits their practical applicability in real-world agricultural settings. This study presents a field-deployable, AI-assisted computer vision system designed for objective, real-time classification of watermelon ripeness. The proposed prototype combines controlled illumination with RGB imaging and convolutional neural networks trained on thousands of annotated outdoor images collected over multiple growing seasons. A phased development strategy—encompassing proof-of-concept modelling, field integration, and multi-season validation—supports robustness against variable lighting conditions and environmental influences. The anticipated outcome is a reliable, non-destructive decision-support tool for growers, capable of identifying ripe fruit for manual harvesting while providing a technological foundation for future autonomous harvesting and precision agriculture applications.
Soybean diseases caused by fungal, bacterial, and viral pathogens represent a major constraint to global agricultural productivity. Although molecular phylogenetic analyses have advanced the understanding of pathogen evolution, the extent to which disease phenotypes reflect evolutionary relationships remains poorly understood. In this study, we developed an integrative framework combining deep learning-based phenotypic analysis with phylogenetic inference to investigate the relationship between soybean disease symptoms and pathogen evolution. An EfficientNet-B0 convolutional neural network (CNN) was trained to classify 10 soybean disease classes comprising 703 leaf images and achieved a mean cross-validation accuracy of 98.72 ± 1.17%, a weighted F1-score of 98.74 ± 1.16%, and a macro F1-score of 98.47 ± 1.74%. Evaluation on a held-out test set generated through image-level partitioning yielded an accuracy of 96.19%, a weighted F1-score of 96.28%, and a macro F1-score of 95.86%. Latent feature embeddings revealed a structured phenotypic space with clear separation among most disease classes and enabled quantitative analyses of phenotypic similarity. To provide biological context, taxonomy-derived distance matrices and sequence-based phylogenetic analyses of the fungal subset using 28S rRNA sequences were compared with CNN-derived phenotypic representations. A Mantel test identified a moderate and statistically significant association between phenotypic and phylogenetic distances (Spearman r = 0.3393, p = 0.0050), indicating that pathogen evolutionary history contributes to disease phenotype while explaining only part of the observed phenotypic variation. Overall, the results demonstrate that deep learning effectively captures biologically meaningful phenotypic information while highlighting that disease symptoms arise from the combined influence of pathogen evolution, host responses, and environmental conditions. This study provides an integrative framework for combining image-based phenotyping with phylogenetic analysis to support biologically informed interpretation of plant disease phenotypes.
Grain quality is defined as the suitability of grain for a particular use. It is usually designated by chemical composition or physical properties of the grain. The ability to measure grain quality is important for identity preservation of specialty grain market classes, for development of new varieties with improved quality through breeding, and for basic scientific studies on the genetic or biochemical control of grain quality traits. This review introduces official methods for measuring maize compositional traits, including protein, starch, oil, amino acid, phytate, and phosphorus content. Additionally, we discuss two nonofficial methods: measuring phytate and available phosphorus levels, and assessing amino acid balance. Phytate and available phosphorous impact the mineral nutrition of grain, while amino acid balance reflects the value of grain as a protein source and the bioavailability of protein. We also describe the use of near-infrared spectroscopy (NIRS) to assess levels of various compounds in maize. NIRS relies on the fact that compounds with differing molecular properties uniquely interact with the near-infrared region (750-2500 nm) of the electromagnetic radiation spectrum, and thus, generate spectral information that can be used to develop calibration models/equations for predicting the concentration of the compounds in grain samples. We discuss how sensitivity, accuracy, precision, throughput, and cost influence the choice of assay used to assess grain quality. Furthermore, we discuss how appropriate experimental design and data analysis can improve analytical outcomes when assessing grain quality.
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
Abstract Leaf color is an important trait affecting vegetable quality, yield, and market value. However, traditional methods for leaf color assessment are often subjective or destructive, which limits accurate and high-throughput phenotyping. In this study, an unmanned aerial vehicle (UAV)–based multispectral imaging platform was used to collect phenotypic data from 214 Chinese cabbage inbred lines at the rosette stage. A multispectral UNet model was applied to segment individual plants, and a membership function was used to quantify leaf color on a continuous scale. Based on these high-throughput phenotypic data, a genome-wide association study was used to identify two candidate genes, BrEMB976 and BrGSH2, on chromosome A06. Subsequent virus-induced gene silencing analysis showed that silencing these genes altered leaf color. In addition, a deep learning-based genomic selection model, BrDeepGS, was developed for leaf color prediction, which achieved a Pearson correlation coefficient of 0.853. These results demonstrate the potential of integrating UAV-based high-throughput phenotyping, candidate gene analysis, and genomic prediction for leaf color evaluation and selection in Chinese cabbage breeding.
As a critical structural component that connects almost all other types of plant organs, the stem system not only supports the weight of the total plant, but also serves as a vital channel for nutriment transportation. Accurate phenotypic measurement of stem instances is of practical significance for assessing crop growth dynamics and predicting yield. To address current 3D phenotyping challenges of crops such as the difficulty in separating stem segments from the stem system and the low accuracy in stem length measurement, we propose a Hierarchical-annotation-guided Phenotypic Framework (HagPF) for Stem Instance Segmentation and Length Measurement in plant point clouds. Specifically, a hierarchical leaf-stem organ instance annotation strategy is devised to effectively train a PSegNet network for leaf and stem instance segmentation. The segmentation is then followed by a shape-adaptive measurement algorithm to automatically measure the length of stem segments that are morphologically diverse in space. On a 3D dataset comprising four crop species, the proposed framework achieved an Intersection over Union (IoU) of 95.45% for organ semantic segmentation and a Mean Weighted Coverage (mWCov) of 87.87% for instance segmentation (both stem and leaf). Regarding to the stem length measurement, the method obtained an average Root Mean Square Error (RMSE) of 1.044 cm and a relative error of 11.907%, outperforming 7 mainstream methods. The relevant dataset and source code can be found at: https://github.com/Jinx00/stem-length-measurement .
Finger millet is an important cereal crop widely cultivated worldwide for food and fodder. Breeding programs aim to select genotypes with desirable architectural traits to develop new varieties with higher yields. In this effort, accurate high-throughput plant phenotyping is essential for accelerating crop improvement. To overcome the time-consuming and labor-intensive process of manual measurements, this study presents a comprehensive 3D imaging pipeline that leverages neural radiance fields (NeRF), 3D gaussian splatting (3DGS), and its advanced extensions (e.g., Feature 3DGS and Gaussian Grouping) to reconstruct, segment, and analyze finger millet yield component traits using multi-view 2D images. First, multiple-view RGB images of a single finger millet plant were captured, and COLMAP was then utilized to estimate the camera poses of the images and reconstruct the sparse point cloud, followed by advanced 3D reconstruction through 3DGS and NeRF. Second, feature 3DGS and gaussian grouping models were used to generate the 3D gaussian representation of finger millet panicles. This single-process framework enabled the generation of high-fidelity 3D point clouds and semantic feature fields without the need for expensive depth sensors or manual annotations. Our results demonstrated the effectiveness of these models in capturing morphological variations across different panicle phenotypes, including compact versus open panicle architectures. In addition, the 3D point clouds of the panicles were utilized to extract structural traits for yield prediction, achieving biologically meaningful correlations with grain productivity. This work highlights the potential of 3DGS-based phenotyping pipelines as a low-cost, near real-time, photorealistic solution for trait quantification, segmentation, and yield estimation in real-world agricultural settings.
Advancements in stay-green phenotyping are increasingly utilizing hyperspectral sensing technology to assess crop response under extreme environmental conditions. Yet, the effectiveness of different spectral features in explaining stay green remains to be fully elucidated. This includes identifying which bands and spectral indices are more effective in capturing the genotypic differences in stay-green traits. The main objective of this study was to evaluate hyperspectral leaf reflectance as a means to estimate stay-green visual scores (SGVS) as an indicator of drought tolerance and to further understand whether chlorophyll absorption-band spectral indices can differentiate SGVS classifications during post-flowering stages of maize. The experiment was conducted over two growing seasons in Germany, comprising 18 maize genotypes under two contrasting water availability conditions. We measured leaf hyperspectral reflectance using a spectroradiometer in the second, fourth, and sixth week after flowering, along with stay-green traits measurements. We employed raw spectral reflectance, hyperspectral vegetation indices (VIs) in combination with random forest (RF) and ANN models to predict SGVS. Results showed that drought stress significantly affected stay-green-related traits and led to a 43.5% decrease in grain yield in the inbred lines. The grain dry yield (GDY) was positively correlated with stay-green visual scores (SGVS), with higher SGVS associated with higher GDY. Stay-green traits were correlated with various VIs, with the best correlation observed for the Chl_NDI (r = 0.91). Stay-green groups were successfully classified using the selected VIs, with the water-absorption band VIs performing better than the chlorophyll-absorption band VIs and other VIs. Similarly, for predicting the SGVS, the water absorption band indices (R² = 0.79 ± 0.04 and RMSE = 0.12 ± 0.01) outperformed the chlorophyll absorption band indices when using RF. Leave-one-out-location/year cross-validation revealed pronounced variation in model transferability driven by environmental and temporal domain shifts. RF consistently outperformed ANN, showing greater robustness to inter-site heterogeneity and interannual variability, whereas performance degraded most in spectrally distinct environments or atypical seasons. Interestingly, RDIS_3b (1280, 1250, 1180 nm), NDIS_2b (2190, 1510 nm), and NDWI2 (860, 1241 nm) were identified as the most critical predictors in the RF models, across merged and separated datasets. These findings demonstrate the potential of spectral signatures, particularly water-absorption band spectral indices, for quantitative phenotyping of stay-green as a proxy for drought tolerance in maize breeding programs; however, multisite, multiyear calibration is needed to enhance generalizability.
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.
• 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.
Rice panicle number per unit area is a key determinant of yield, but manual counting remains time-consuming and labor-intensive. This study proposes CEG-YOLO, a lightweight deep learning model for in-field rice panicle detection and counting using consumer-grade RGB imaging devices. The model introduces three improvements to YOLOv11s to address specific challenges in field scenarios: C2f-Fast replaces standard convolutions with depthwise convolutions to reduce computational cost for edge deployment; SPPF-ECA integrates an attention mechanism to suppress complex background interference; and GhostConv reduces feature redundancy to improve detection of dense panicles. A dataset of 5,175 images was collected from four rice cultivars planted at three densities using an iPhone 12. The proposed model achieved 93.9% average precision (AP) on the test set, outperforming YOLOv11s which achieved 89.1%, while reducing parameters to 7.8 million and floating-point operations (FLOPs) to 16.5 billion. Robustness evaluation yielded coefficients of determination (R²) values of 0.91 to 0.94 for lighting, 0.89 to 0.94 for planting density, and 0.90 to 0.94 for cultivar. A cross-year field test using an NVIDIA Jetson Orin NX edge device on 120 quadrats in 2025 achieved R² of 0.91, root mean square error (RMSE) of 4.0, and mean absolute error (MAE) of 3.3 at 20 frames per second, confirming practical deployability. This study demonstrates that smartphone-based proximal phenotyping with an optimized deep learning model can provide accurate, low-cost rice panicle counting for breeding and production applications.
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.
Wild rice (Oryza spp.) harbors abundant genetic variation and represents an important germplasm resource for improving yield-related traits in cultivated rice. Panicle number is a key phenotypic trait for evaluating tillering capacity and yield potential in wild rice. However, existing approaches for acquiring panicle-number phenotypes remain limited by low efficiency, high dependence on manual operation, and cumbersome matching between plant targets and accession identifiers. In this study, we proposed an AI-driven UAV phenotyping framework for high-throughput panicle counting of wild rice accessions for germplasm evaluation. The framework integrates field plant localization, accession identifier binding, flight route planning, plant-by-plant video acquisition, spatiotemporal registration, video slicing, and panicle detection and tracking, enabling structured panicle-number outputs indexed by accession identifier. To address the small scale, loose structure, morphological variation, and wind-induced swaying of wild rice panicles in UAV imagery, a wild rice panicle detection model was constructed, and WRPD-Tracker was developed for cross-frame identity association and non-redundant counting. The wild rice panicle detection model achieved an AP@50 of 91.56%, representing a 6.16-percentage-point improvement over the DEIM baseline, with 3.70 M parameters and 6.55 G FLOPs, while WRPD-Tracker achieved a HOTA of 65.1% and a MOTA of 79.0%, representing a 5.8-percentage-point improvement in HOTA over the baseline tracker. At the final counting level, UAV-based counts were highly consistent with manual ground counts, with an R² of 0.992 and an MAE of 0.37 panicles. This framework enables batch acquisition of panicle-number phenotypes in wild rice and provides quantitative support for germplasm evaluation, panicle-number trait comparison, and subsequent yield-related phenotypic studies.
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.
Plant diseases have continued to threaten agricultural productivity, while manual inspection methods have remained inefficient and prone to subjectivity. This study proposed and assessed a hybrid framework integrating a Convolutional Neural Network (CNN) with a Large Language Model (LLM) to perform image-based plant leaf disease classification accompanied by interpretable diagnostic explanations. EfficientNetV2-M was employed as the visual backbone and trained on 11 selected classes of apple, grape, and potato leaf images derived from the PlantVillage dataset. A structured data splitting strategy was applied to ensure reliable model validation and unbiased testing. The classification capability of the CNN component was examined through standard multi-class evaluation indicators, including class-wise predictive consistency and error distribution analysis. Experimental results indicated that the model delivered highly consistent predictions, reaching a peak test accuracy of 99.79%, reflecting its robustness in distinguishing visually similar disease patterns. To overcome the black-box limitation, prediction outputs were transformed into structured prompts and processed by GPT-4o to generate contextual explanations. The generated narratives systematically described observable symptoms, highlighted distinguishing characteristics, and suggested initial management actions. Overall, the proposed hybrid system demonstrated that combining high-performance visual recognition with language-based reasoning enhanced both diagnostic reliability and interpretability in digital agriculture applications.
Reproduction assets foundThe paper uses the public PlantVillage color dataset (11 apple/grape/potato classes, 9,385 images) and explicitly points to it in the DATA AVAILABILITY statement as the replication package data. The authors also provide a public Streamlit demonstration of their hybrid CNN–LLM system. No author analysis code or trained-Dataset · publicaper. The research was conducted for academic purposes, and no financial, commercial, or
personal relationships influenced the study design, data analysis, interpretation of results, or
preparation of the manuscript.
DATA AVAILABILITY
The data associated with this study are publicly available online in the replication package.
[https://github.com/spMohanty/PlantVillage-Dataset/tree/master/raw/color]
AUTHOR CONTRIBUTIONS
Frenky Riski Gilang Pratama: Conceptualization; Programming and coding
implementation; Methodology; Writing-Original Draft. Sugiarto Surono: Conceptualization;
Methodology; Supervision; Writing-Review & Editing. Aris Thobirin: Proofreading Paper;
Writing-Review & Editing; FunOpen asset ↗https://github.com/spMohanty/PlantVillage-Dataset/tree/master/raw/colorpdf-raw-page:10 lines:1-52Plant phenotyping relevance matchCrossref · OpenAlex · checked 14 Sept 2026
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.
【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.
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.
Image-based techniques are widely used in plant disease classification to support agricultural productivity and facilitate early detection. This study presents a comparative analysis between a baseline convolutional neural network (CNN) and a ResNet50-based transfer learning model for tomato leaf disease classification. A publicly available dataset containing five categories—tomato bacterial spot, tomato late blight, tomato septoria leaf spot, tomato yellow leaf curl virus, and healthy leaves—was used in the experiments. Model performance was evaluated using several standard classification metrics, with emphasis on overall accuracy. The baseline CNN achieved an accuracy of 97.0%, whereas the ResNet50 model reached 99.6%. The results demonstrate that the ResNet50 model produces more stable and reliable predictions, particularly when distinguishing between visually similar disease classes. These findings confirm that transfer learning can effectively improve classification performance in plant disease recognition tasks.
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.
Early screening of tomato leaf diseases is important because foliar symptoms can reduce plant vigor and delay appropriate crop management. This study develops an Android-based tomato leaf disease classification system using MobileNetV2 as a lightweight convolutional neural network (CNN) architecture. The contribution of this work is the integration of model training, independent testing, and on-device Android deployment that supports camera and gallery inputs without relying on server-side computation. The dataset consisted of 1,200 balanced tomato leaf images from five classes: bacterial spot, late blight, target spot, tomato yellow leaf curl virus, and healthy leaf. Images were resized, normalized, augmented for training, and divided into training, validation, and independent testing subsets. The model obtained 94.12% training accuracy, 93.00% validation accuracy, and 89.00% independent test accuracy. The confusion matrix showed that tomato yellow leaf curl virus was classified without error, whereas bacterial spot, late blight, target spot, and healthy leaves produced several misclassifications because of similar lesion and discoloration patterns. The results show that MobileNetV2 is suitable for lightweight mobile disease screening, although larger field datasets, cross-validation, model comparison, and explainability analysis are still needed for broader deployment.
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.
Crop residues support soil health by reducing erosion, improving water retention, and contributing to carbon sequestration. Accurate estimation of crop residue biomass is essential for understanding residue distribution patterns and improving sustainable land management practices. Remote sensing, especially high-resolution UAV-based imaging, is a powerful tool for monitoring residue over agricultural fields, and many studies use remote sensing datasets for mapping residue cover (a 2D metric). However, few studies have evaluated residue biomass using remote sensing, despite biomass being more ecologically informative. This study uses high-resolution UAV multispectral imagery to predict crop residue biomass using feature selection and machine learning. Candidate predictors included raw bands, spectral indices, texture metrics, and digital-surface-model-derived topographic variables. Three feature selection methods—recursive feature elimination with cross-validation, Pearson correlation screening, and least absolute shrinkage and selection operator regression, were applied on the training set to identify informative predictors. Four machine learning models (Random Forest Regression, Support Vector Regression, CatBoost, and k-Nearest Neighbors [kNN]) were evaluated individually and in combination using simple averaging, weighted averaging, and stacked ensemble strategies. Results show that Pearson-selected features paired with kNN achieved the best performance (R 2 = 0.61, RMSE = 188.71 g m -2 ). Ensemble approaches did not outperform the best individual model, suggesting limited benefit from meta-learning under small-sample conditions. Across selection methods, red- and blue-band-related predictors were consistently retained, while textural and topographic variables were selected more selectively, indicating context-dependent contributions. Overall, simpler models with targeted feature selection can outperform more complex ensembles for UAV-based crop residue biomass estimation.
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.
Sugarcane plant diseases pose a significant threat to agricultural productivity, yet early and accurate identification remains challenging for farmers due to the limitations of manual inspection. This study proposes a sugarcane leaf disease classification system using ConvNeXt V2 Tiny, a modern convolutional architecture with a Global Response Normalization (GRN) mechanism, combined with an ensemble Stratified K-Fold Cross Validation strategy (K=6) to improve generalization on real-world field data. A dataset of 2,948 leaf images spanning five classes (Red Rot, Mosaic, Rust, Yellow Leaf, and Healthy) was used, with field-collected images held out as a fixed test set. The ensemble model achieved a mean validation accuracy of 98.49% ± 0.58% across six folds and a test accuracy of 98.39% on 427 unseen field images, with macro-average precision, recall, and F1-score each reaching 98%. ConvNeXt V2 Tiny substantially outperformed ResNet-50 (87.35%) and EfficientNetV2-S (83.37%) under identical experimental settings, demonstrating superior generalization across the domain gap between curated and field data. The primary contribution of this study is the first application of ConvNeXt V2 Tiny with ensemble K-Fold strategy for sugarcane disease classification, offering high accuracy with moderate computational complexity (28.6M parameters) and practical deployability, as demonstrated through the SugarScan web application.
Reproduction assets foundThe paper's phenotyping inputs include a public Kaggle dataset (Sugarcane Leaf Disease Dataset, SLD) of sugarcane leaf disease images used for training/validation, plus field-collected images. Only the Kaggle dataset qualifies as a paper-specific public asset with an authors' URL; no author analysis code, trained modelDataset · publicsecondary data from the Sugarcane Leaf Disease Dataset (SLD)
available publicly on Kaggle (https://www.kaggle.com/datasets/pritpal2873/sug arcane-leaf-disease-dataset)Open asset ↗Kaggle · pritpal2873/sugpdf-page:2 lines:54-60Plant phenotyping relevance matchCrossref · checked 14 Sept 2026
Published31 Jul 2026World Journal of Advanced Engineering Technology and SciencesCited by 0 · OpenAlex ↗
Proper diagnosis of plant disease and nutrient deficiency is crucial in enhancing crop productivity, reducing yield losses and early agricultural interventions. This paper introduces the machine learning approach for automated identification of plant diseases and their severity by analyzing the images of plant leaves. The proposed framework involves image preprocessing, feature extraction, classification and severity estimation, which would enable the accurate identification of various types of disease and estimation of their severity levels while detecting nutrient deficiencies. An extensive image dataset of healthy, diseased and nutrient deficient leaves was used to train and test the models. As illustrated by the experimental results, the proposed framework outperforms the existing machine learning and deep learning methods for plant disease identification and nutrient deficiency detection with the classification accuracy of 98.76% and 98.14% respectively. In addition, the severity assessment module estimates well to enable accurate pesticide and nutrient application, which minimizes chemical use. The proposed framework provides a scalable, efficient, and precise approach for smart crop health monitoring and precision agriculture applications.
Broccoli is a globally significant vegetable, but climate change and soil salinization increasingly threaten its productivity. Precise seedling phenotyping is essential for selecting salt-tolerant germplasm, yet traditional manual methods are labor-intensive and error-prone. This study develops LBD-PointNet++, an optimized 3D point cloud semantic segmentation model for automated phenotypic parameter extraction of broccoli seedlings at the germination and early developmental phase under salt stress. High-fidelity 3D point clouds were reconstructed from a precision three-view imaging system using Structure from Motion (SfM) algorithms. LBD-PointNet++ introduces three core optimizations: (1) a Large Kernel Attention (LKA) mechanism using 3D sparse decomposition to capture long-range global dependencies; (2) a Dual Uncertainty and Shape-Adaptive Sampling (DUSAS) mechanism to preserve high-frequency features of fragile stems and margins; and (3) a joint Boundary-Aware Nested Contrastive and Adaptive Varifocal Joint Loss (BNCV-Loss) to effectively isolate overlapping leaves. Experimental results demonstrate superior performance, achieving an overall mean Intersection over Union (mIoU) of 88.07% across all three categories (Leaf, Stem, and Pot) and a Mean F1-score of 93.48%. Compared to state-of-the-art Transformer architectures like PTv3, LBD-PointNet++ achieves higher accuracy with less than 6% of the parameter volume and over twofold faster inference speed. Furthermore, dynamic monitoring across NaCl gradients (0-250 mmol/L) revealed a potential non-linear threshold effect, identifying 100 mmol/L as a preliminary phenotypic threshold under these conditions. Beyond this threshold, growth inhibition intensified rapidly; At 250 mmol/L, plant height decreased by 54.43% and the 3D entity volume shrank to approximately one-fifth of the control group. In summary, LBD-PointNet++ provides a high-efficiency solution for phenotypic identification and digital breeding of salt-tolerant Brassicaceae crops.
Common beanGrapevineLeafClassificationObject detectionStress / disease detectionDisease symptoms / severity
Plant leaf disease is a grave risk to crop suitability and agricultural sustainability, and therefore, the ability to make early and accurate diagnosis is a mandatory need in the contemporary precision farming system.In recent years, deep learning has gained significant attention for image-based plant disease detection.Despite its effectiveness, model performance can be influenced by factors such as redundant feature representations and sensitivity to hyperparameter selection.To address these challenges, this study proposes a nested hybrid optimization framework that combines the Cuckoo Search Algorithm (CSA) for channel selection with the Beluga Whale Optimization Mechanism (BWOM) for tuning the hyperparameters of a Convolutional Neural Network (CNN).The proposed approach is independently evaluated on bean and grape leaf datasets under consistent experimental conditions to assess its disease classification performance.In addition to strong predictive performance, the framework incorporates explainable AI (XAI) techniques, namely Gradient-weighted Class Activation Mapping (Grad-CAM) and Gradientweighted Class Activation Mapping Plus Plus (Grad-CAM++), to enhance model interpretability.These approaches highlight the most significant visual features influencing predictions, thereby providing valuable insights for agronomists and fostering trust in AI-based systems.Experimental results show that the proposed CSA-BWOM optimized CNN achieves classification accuracies of 99.61% and 99.38% on the bean and grape datasets, respectively, outperforming baseline CNN models and exhibiting competitive performance when compared to a few existing approaches.
Identifying the complex leaf disease is a major challenge in crop production. Traditional Capsule Networks face challenges in capturing the complex features associated with various plant diseases, even though they can effectively retain spatial hierarchies. To overcome challenges, this research presents a novel Capsule Networks framework for the identification of leaf disease and its classification. The research proposes a Feedback-Enabled Capsule Routing and a Capsule-Level Attention mechanism. The Feedback-Enabled Capsule Routing method that refines classification outcomes also introduces top-down contextual signals, whereas the Capsule-Level Attention mechanism dynamically weighs capsule outputs based on their relevance. Incorporating this, we also integrated Primary Capsule Squashing and Dynamic Routing. All methods are grounded in formal mathematical definitions. The XAI techniques are integrated to visualize the model's decision-making process for transparency. Empirical evaluations demonstrate the model's effectiveness in achieving classification accuracies of 98.97% and 98.49% in the PlantVillage and the PLD datasets, respectively. This framework addresses the trust gap in AI-driven agricultural diagnostics, allowing for a transparent and reliable method of disease detection in agricultural applications.
Soybean pod-related traits are important for seed development, yield formation, and cultivar evaluation. However, conventional measurements are inefficient and have limited ability to characterize complex pod features such as curvature, local enlargement, and continuous color variation. In this study, mature pod images of 187 cultivated soybean accessions collected across two successive years were analyzed using a deep learning-assisted image phenotyping approach. Eight pod-related traits related to size, morphology, and color were extracted from pod images. A genome-wide association study (GWAS) was performed using 61,541 high-quality SNP markers to dissect the genetic basis of these image-derived pod traits. A total of 16 stable loci associated with pod size, morphology, and color traits were identified across 11 chromosomes. Among these loci, eight were not reported in the previous image-based soybean pod GWAS study. Based on SoyBase gene annotation and Gene Ontology biological process information, 32 biologically relevant candidate gene records were prioritized within the corresponding candidate genomic intervals, while pod-related expression profiles and SoyBase association information were used as supporting evidence for candidate gene evaluation. These findings indicate that refined image-derived traits can provide complementary genetic information beyond conventional pod measurements and offer additional opportunities for dissecting soybean pod development, morphology, and mature pod color variation.
Agriculture plays a crucial role in the Indian economy. Early detection of plant diseases is very much essential to prevent crop loss and further spread of diseases. Most plants such as apple, tomato, cherry, grapes show visible symptoms of the disease on the leaf. These visible patterns can be identified to correctly predict the disease and take early actions to prevent it. This can be overcome by the use of machine learning and deep learning algorithms. Hence, we are proposing a method that which is detecting the disease of a tomato plant from their leaf images. Here the process is performed with the deep learning algorithms Convolutional Neural Network (CNN), and MobileNet which is a one of the transfer learning method of CNN. Once after training the dataset with the algorithms, the accuracy of algorithms is compared and the images are classified. And the precautions are also provided for the classified plant.
Penyakit Basal Stem Rot (BSR) merupakan penyakit penting pada tanaman kelapa sawit yang dapat menurunkan produktivitas. Analisis berbasis citra dapat mendukung deteksi penyakit, tetapi karakteristik distribusi warna perlu dipahami sebelum tahap klasifikasi. Penelitian ini bertujuan menganalisis distribusi warna citra daun kelapa sawit kelas Healthy dan BSR menggunakan Histogram pada ruang warna RGB dan HSV serta mengevaluasi kemampuan diskriminatif fitur warna. Dataset terdiri atas 2.438 citra dari repositori terbuka Roboflow. Setiap citra diproses melalui resize 224×224 piksel, Gaussian Blur, konversi RGB ke HSV, serta ekstraksi 12 parameter statistik berupa mean dan standar deviasi. Perbedaan distribusi antar kelas diuji menggunakan Mann–Whitney U dan besarnya perbedaan dihitung menggunakan Cohen's d. Kemampuan diskriminatif G_Mean, S_Mean, dan V_Mean dievaluasi menggunakan Receiver Operating Characteristic (ROC) dan Area Under the Curve (AUC). Hasil menunjukkan ketiga fitur memiliki perbedaan signifikan (p-value = 0,0000) dengan ukuran efek besar. Nilai AUC G_Mean, S_Mean, dan V_Mean masing-masing sebesar 0,9794, 0,7362, dan 0,9790. Hasil menunjukkan bahwa analisis distribusi warna dapat mengidentifikasi fitur diskriminatif sebagai dasar pemilihan fitur dan perancangan pra-pemrosesan citra BSR.
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 matchCrossref · checked 15 Sept 2026
Published30 Jul 2026JOIV : International Journal on Informatics VisualizationCited by 0 · OpenAlex ↗
The prompt identification and precise categorization of chili plant diseases are crucial for promoting sustainable agriculture and reducing crop losses due to pests and pathogens. This research presents an automated detection method using the Single Shot Detector (SSD) framework, with ResNet-50 as the backbone and a Feature Pyramid Network (FPN) to manage multi-scale feature representations. Dataset including 14,248 images of chili plants, categorized into six classes like Leaf Spot, Rotten Fruit, Healthy Fruit, Healthy Leaf, Mosaic Curl, and Yellowing, underwent preprocessing involving segmentation, resizing, and augmentation, followed by a division into 90% training data and 10% testing data. Transfer learning was implemented using a COCO-pretrained SSD ResNet-50 FPN model, enhanced with cosine-decay learning-rate scheduling and momentum optimization. The assessment results indicated an overall accuracy of 92.7%, with the highest F1-scores achieved for Healthy Fruit (0.965) and Rotten Fruit (0.967). Under the COCO evaluation protocol, the model achieved an mAP@0.5 of 91.5% and mAP@[0.5:0.95] of 65.2%. Model ran at approximately 30 FPS on an NVIDIA T4 GPU. Reduced precision values were noted for Leaf Spot (0.866) and Mosaic Curl (0.850), suggesting a propensity for misclassification due to visual similarities among disease symptoms. Nonetheless, all classes attained F1-scores exceeding 0.86, illustrating the robustness of the proposed model. Importantly, the SSD-ResNet-50 approach offers both efficiency and accuracy within a single pipeline, enabling rapid inference practical for real-world applications. These findings emphasize the potential of deep learning-based solutions to strengthen plant disease monitoring systems. In conclusion, SSD with ResNet-50 FPN provides an effective and scalable methodology for the automated identification of chili plant diseases, contributing directly to sustainable agriculture and improved crop management practices.
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.
Camera-based visual sensing provides a non-destructive and scalable approach for monitoring strawberry diseases and pests in greenhouse environments. However, greenhouse images acquired under practical cultivation conditions often contain early-stage tiny lesions, complex leaf backgrounds, uneven target scales, illumination variations, and partial occlusions, making accurate and efficient visual detection challenging. To address these issues, this study proposes YOLOv8n-DSLW (YOLOv8n enhanced by Dense reuse, Shuffle attention, LSKA-LAMP lightweight modeling, and Wise-IoU optimization), an AI-enabled vision-sensing detection model based on YOLOv8n for tiny strawberry disease and pest detection. Specifically, Shrink Residual Dense Block (ShrinkRDB) dense connection blocks and the C2f with Shuffle Attention (C2fSA) module are introduced to preserve weak lesion textures and suppress background interference in greenhouse visual data. A high-resolution P2 detection layer combined with Wise-IoU (WioU) dynamic regression loss is further incorporated to enhance tiny-target perception and localization. In addition, the Spatial Pyramid Pooling-Fast with Large Separable Kernel Attention (SPPF-LSKA) module strengthens contextual modeling under occlusion and clutter, while Layer-Adaptive Magnitude-based Pruning (LAMP) is adopted to mitigate model redundancy and improve the accuracy-efficiency balance. Experiments on a self-collected greenhouse strawberry disease and pest dataset show that YOLOv8n-DSLW achieves a mean Average Precision at 0.5 IoU threshold (mAP@0.5) of 94.3% and a mAP@0.5:0.95 of 77.5%, outperforming the YOLOv8n baseline. The final model has a parameter count of 4.386 M and a computational cost of 27.6 GFLOPs, achieving a frame rate of 45 FPS on the test workstation. It shows application potential for real-time visual monitoring in greenhouses under controlled data acquisition conditions. The results demonstrate that the proposed method improves tiny lesion detection under dense targets, complex backgrounds, and leaf occlusions, providing an AI-enabled vision-sensing framework for automated strawberry health monitoring in greenhouses. Nevertheless, due to limitations associated with imaging equipment, dataset representativeness, and the inherent constraints of the algorithm, further optimization and validation are required to support large-scale field deployment.
Reproduction assets foundThe paper's self-collected greenhouse strawberry disease/pest image dataset (with COCO annotations and train/test splits) is explicitly stated as publicly deposited on GitHub at the allowed URL. No author analysis code or trained model checkpoint is mentioned as publicly available.Dataset · publicThe dataset used in this study, including the training and independent test subsets, has been uploaded to a GitHub repository for dataset verification and is available at: https://github.com/dataset-review-2026/strawberry-dataset (accessed on 26 July 2026).Open asset ↗dataset-review-2026/strawberry-datasetlines:111-131Plant phenotyping relevance matchOpenAlex · checked 14 Sept 2026
Agriculture continues to be one of the principal contributors to the economy and food security of developing nations, yet farmers regularly face difficulties such as unpredictable weather, variable soil conditions, and crop diseases that reduce quality and income. This paper presents a Crop Prediction and Plant Disease Detection System that integrates machine learning and deep learning techniques within a single web-based platform. The system accepts agricultural parameters — nitrogen, phosphorus, potassium, temperature, humidity, pH, and rainfall — and applies a Random Forest regression model to estimate the expected crop. In parallel, it allows farmers to upload images of crop leaves, which are pre-processed and classified by a Convolutional Neural Network (CNN) to identify plant diseases and recommend suitable treatment. The application is built using ASP.NET Core for the user-facing interface, authentication, and dashboard, while a Python-based REST API hosts the machine learning and deep learning models; Microsoft SQL Server is used for persistent storage of user data, predictions, and disease records. The proposed system combines two traditionally separate functions — crop estimation and disease diagnosis — together with treatment recommendations and prediction history, into a single decision-support tool for precision agriculture, and was validated through unit, integration, and system-level testing
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.
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.
Plant phenotyping plays a critical role in understanding plant health and improving agricultural productivity by enabling quantitative analysis of disease-related physiological characteristics.Among these, leaf diseases significantly impact crop yield and quality, necessitating accurate and automated phenotyping approaches.Traditional phenotyping methods rely on manual inspection or handcrafted feature extraction, which are time-consuming, prone to human error, and lack scalability under diverse environmental conditions.This study introduces a deep learning (DL)-based approach for image-based plant phenotyping, focusing on the classification of disease-affected traits.The developed method integrates segmentation-based region extraction, dual-branch feature learning, and attention-based feature fusion.Initially, input images are preprocessed and passed through a TransUNet-based segmentation module to isolate phenotypically relevant leaf regions while suppressing background interference.Both the original image and the segmented region are then processed using a RegNet-based feature extraction network to capture global structural information and localized disease-specific characteristics.The extracted features are fused using an attention-based mechanism, followed by fully connected layers for multiclass classification.Experimental results obtained on the controlled PlantVillage grape leaf dataset, which serves as a standardized benchmark for plant disease classification, demonstrate an overall classification accuracy of 97.8%, with precision, recall, and F1-score values of 97.7%, 97.9%, and 97.8%, respectively.In addition, the segmentation module achieves an Intersection over Union (IoU) of 94.1% and a Dice score of 96.8%, confirming its effectiveness in isolating relevant phenotypic regions.
Sebastian Tonn · Mon-Ray Shao · Jos de Wit · Rami Mousa · Iñigo Bañales · Roy A M van Beekveld · Arjen N Bader · Henriëtte D L M van Eekelen · Ric C H de Vos · Eefjan Breukink · Guido Van den Ackerveken
Abstract Effective plant disease phenotyping is crucial for resistance breeding, but traditional visual assessment is often inaccurate and inefficient. This is particularly challenging when breeding lettuce (Lactuca sativa) for resistance to downy mildew, given the obligate biotrophic lifestyle of the causal pathogen Bremia lactucae. We discovered that B. lactucae-infected lettuce exhibits patches of increased blue-green fluorescence (BGF) under UV-A excitation from 6 d post-inoculation, preceding visible symptoms. Co-localization of BGF with hyphae, visualized with trypan blue, indicates that BGF is induced by downy mildew colonization. We therefore investigated its potential for non-invasive disease detection and quantification, as well as the underlying physiological changes. Using a custom imaging system, we demonstrate that BGF leaf area correlates with downy mildew severity and can be automatically quantified via a U-Net-based convolutional neural network, enabling early, objective disease assessment. Exploring transcriptomic and metabolomic changes associated with BGF, we found that induction of the phenylpropanoid pathway led to accumulation of caffeoylquinic acids, whose fluorescence spectra overlap with that of BGF tissue, supporting the hypothesis that these compounds contribute to the fluorescence signal. BGF imaging offers a powerful tool for phenotyping in lettuce breeding and for identifying quantitative resistance traits that support durable downy mildew resistance.
Agriculture plays a pivotal role in ensuring global food security, economic stability, and sustainable development. Plant diseases significantly reduce agricultural productivity, resulting in substantial economic losses and threatening food supply worldwide. Early and accurate identification of plant leaf diseases enables timely intervention, minimizes crop damage, and enhances agricultural yield. Traditional disease diagnosis relies heavily on visual inspection by agricultural experts, making the process labor-intensive, subjective, and unsuitable for large-scale deployment. Recent advances in artificial intelligence, particularly deep learning, have transformed plant disease diagnosis by enabling automatic feature extraction and highly accurate image-based classification. This review presents a comprehensive analysis of recent developments in deep learning techniques for plant leaf disease identification. Various convolutional neural network (CNN) architectures, including AlexNet, VGGNet, ResNet, DenseNet, EfficientNet, MobileNet, Inception, and Xception, are critically reviewed along with modern transformer-based models such as Vision Transformer (ViT), Swin Transformer, and hybrid CNN–Transformer frameworks. The paper also examines transfer learning strategies, object detection methods including YOLO and Faster R-CNN, and semantic segmentation approaches such as U-Net and DeepLabV3+. Publicly available benchmark datasets, including PlantVillage, PlantDoc, AI Challenger, Cassava Leaf Disease, and Rice Leaf Disease datasets, are discussed in terms of dataset diversity, annotation quality, and practical applicability. Furthermore, image preprocessing techniques, data augmentation methods, evaluation metrics, and deployment considerations for mobile and edge devices are comprehensively reviewed. The paper identifies current research challenges, including dataset imbalance, environmental variability, model interpretability, computational complexity, and limited real-world generalization. Finally, emerging research directions such as explainable artificial intelligence, federated learning, multimodal learning, self-supervised learning, lightweight architectures, and edge AI are discussed to provide future research opportunities. This review serves as a valuable resource for researchers, practitioners, and agricultural technologists interested in developing robust, scalable, and intelligent plant disease identification systems.
Enhancing photosynthesis is an important approach to improve crop yields. Photosynthesis, as a key factor determining crop yield, is an important approach to increasing crop production and addressing global food security issues. Improving its efficiency is crucial in this regard. However, traditional photosynthetic phenotyping has long been a bottleneck in crop breeding due to time-consuming data collection. In this study, we simultaneously measured the spectral reflectance and the net photosynthetic rate (Pn) of soybean leaves to develop a high-precision model for estimating Pn based on hyperspectral data. By applying this model, we evaluated Pn in 219 soybean materials. A multi-environment genome-wide association study (GWAS) based on multi-environmental prediction Pn was carried out using the 3VmrMLM method, and 24 significant quantitative trait loci (QTLs) and four suggestive QTLs were identified. Among them, 24 QTLs overlapped with multiple previously reported QTL related to photosynthesis, chlorophyll content, quality, etc., or with genes related to key agronomic traits such as yield. Additionally, four new QTLs were discovered, and four candidate genes potentially associated with Pn were identified. Further, haplotype analysis identified their optimal haplotypes. This study presents a robust and nondestructive hyperspectral model for estimating the photosynthetic rate in soybeans, which is successfully applied to genetic analysis, yielding stable and biologically meaningful results. The approach offers an effective means to explore the genetic basis of photosynthesis and provides a solid theoretical foundation for large-scale, monitoring of soybean photosynthetic physiology.
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.
Introduction Banana production is increasingly threatened by fungal diseases such as Fusarium wilt and Black Sigatoka, posing severe risks to food security and agricultural economies. Recent image-based approaches using deep learning have shown high predictive capacity for plant disease recognition; however, their limited transparency, calibration uncertainty, and sensitivity to domain shifts can restrict their use in decision-support workflows that require auditability. Methods This study proposes an interpretable and calibrated Artificial Intelligence framework for multiclass banana disease-pattern characterization based on radiomic feature analysis of RGB leaf images. Radiomic features were extracted from HSV-segmented banana leaf regions, resulting in a dataset of 14,763 samples characterized by 103 quantitative descriptors and labeled as Healthy, Sigatoka, or Fusarium wilt race 1. Results Among the evaluated radiomics classifiers, the calibrated Random Forest achieved accuracy = 0.85, balanced accuracy = 0.84, macro-F1 = 0.84, and macro ROC-AUC OvR = 0.95 on the held-out test set. Bootstrap analysis yielded 95% confidence intervals of [0.8406, 0.8691] for accuracy and [0.8306, 0.8606] for balanced accuracy. Three deep learning baselines trained on the same partition achieved higher predictive performance: MobileNetV3 with accuracy = 0.96, macro-F1 = 0.95, and macro ROC-AUC OvR = 0.97; EfficientNet with accuracy = 0.96, macro-F1 = 0.97, and macro ROC-AUC OvR = 0.96; and ResNet-18 with accuracy = 0.96, macro-F1 = 0.96, and macro ROC-AUC OvR = 0.96. Discussion The CNNs produced strong classification performance on the evaluated repositories, and the radiomics approach demonstrated to be a complementary interpretable and explainable calibrated reference model. SHAP, LIME, permutation importance, accumulated local effects, calibration curves, and Brier score decomposition supported feature-level inspection of the final model.
Leaf disease diagnosis needs models that are accurate enough for agronomic use yet small enough for constrained computing settings. This study examines a late-attention MobileNetV2 design in which one Convolutional Block Attention Module (CBAM) is inserted between the last MobileNetV2 convolutional map and global average pooling. The experiments use 54,306 controlled-background PlantVillage images spanning 38 classes. Under a uniform saved-model re-evaluation, MobileNetV2 + CBAM obtained 97.17% accuracy and 97.15% weighted F1-score, whereas MobileNetV2 obtained 96.78% and 96.73%. On the converted models, paired testing gave a 0.64-percentage-point accuracy advantage for the CBAM variant (95% CI: 0.31–0.96; exact McNemar p
Plants encounter multiple abiotic stresses. Among them, heat and drought stress play a substantial role in reducing the agricultural productivity of commercial plants. Hence, wild and underutilized plants can be a potential alternative as they are naturally tolerant to extreme climatic conditions and are a rich source of nutrition. Manual stress and disease detection is a laborious and expensive process, and hence automation in this field is required to reduce agricultural losses. This study evaluates the prediction and detection of abiotic stress in Acacia senegal bipinnate leaves, exploring various stress-induced changes using machine learning (ML) algorithms and biochemical analysis. A. senegal , an underutilized edible desert legume, was grown under controlled greenhouse conditions. After 2 months, these plants were segregated into groups and subjected to heat and drought treatments. Image acquisition was performed to obtain a dataset of 3,454 images of A. senegal leaves. Physiological parameters, such as fresh and dry leaf weight, shoot length, number of leaves, and biochemical assays like antioxidant assay (DPPH), total phenolic content (TPC), and total flavonoid content (TFC), were determined. LC-MS/MS analysis was conducted to identify over 50 phytochemical compounds. A hybrid model was developed consisting of a fine-tuned EfficientNet-based Convolutional Neural Network (CNN) followed by a Support Vector Machine (SVM) for the binary classification of A. senegal leaves. The model distinguishes between healthy and stress-affected unhealthy leaves and achieved an accuracy score of 86.6%. This report provides a significant lead toward stress phenotyping and prediction of a bipinnate leaf plant using ML algorithms. The overall study is useful to understand how the stress encountered by arid plants alters the nutritional quality.
Reproduction assets foundThe paper's data availability statement explicitly makes the 3,454-image A. senegal leaf imaging dataset public on Zenodo and the ML implementation source code public on GitHub; both are paper-specific, public, and actionable.Dataset · publicThe plant leaf imaging data used in the work is publicly available at https://doi.org/10.5281/zenodo.16531486.Open asset ↗zenodo · 10.5281/zenodo.16531486html-lines:480-497Code · publicThe source code of the implementation is available at https://github.com/softwareinnovationslabBITS/CDRF_ASenegal_MLImagingOpen asset ↗github · softwareinnovationslabBITS/CDRF_ASenegal_MLImaginghtml-lines:480-497Plant phenotyping relevance matchEurope PMC · checked 14 Sept 2026
Rice leaf diseases pose a major threat to crop productivity and global food security. Therefore, rapid and accurate disease diagnosis is essential for effective crop management. Despite the advancements made by deep learning algorithms in automated disease detection, the majority of current deep learning-based models are built and evaluated on single-source datasets. Therefore, the generalization capability of these models is still questionable. In order to overcome this problem, this paper presents an interpretable deep learning approach for classifying diseases in rice leaves using ConvNeXt-Tiny and Grad-CAM. The proposed semantic class normalization technique is used to harmonize the heterogeneous data classes to form a seven-class classification system. The model performance was analyzed based on two available public datasets for rice leaf disease and a combination of both. After performing five individual runs, the average classification accuracies were calculated to be 99.13±0.37%, 95.28±0.63%, and 97.27±0.41% for Datasets 1, 2, and the combined dataset, respectively. The confusion matrix analysis indicated minor misclassification errors in the form of false positives for the disease classes which have similarities in appearance. The training and validation curves showed consistent learning with minimum overfitting. Moreover, Grad-CAM analysis confirmed the focus of the model on the disease-specific regions. These results collectively show that the combination of transfer learning, semantic class harmonization, and explainable AI is a stable platform to conduct automated rice disease detection. It is evident from these outcomes that reliable performance can be achieved even under heterogeneous imaging conditions. Consequently, there exists great potential for this approach in agricultural decision-support systems.
Plant disease detection is critical for sustainable agriculture and food security. While deep learning models achieve high accuracy in leaf disease classification, their black box nature poses limitations for trust and adoption among agricultural practitioners. This study presents a comparative evaluation of three convolutional neural network architectures (ConvNeXt-Tiny, MobileNetV2, and VGG16) for classifying potato, maize, and pepper leaf diseases, with emphasis on explainability through Gradient-weighted Class Activation Mapping (Grad-CAM). The experimental results demonstrate that ConvNeXt-Tiny achieves 99-100% accuracy across all plant species, MobileNetV2 attains 97-100% accuracy with lower computational requirements, and VGG16 yields 97-99.5% accuracy. Grad-CAM visualizations reveal that modern architectures precisely focus on lesion regions, whereas older models occasionally attend to irrelevant features such as leaf veins and edges. Misclassification analysis identifies shadows and natural leaf patterns as primary error sources. This research demonstrates that explainable artificial intelligence is not merely complementary but essential for developing trustworthy agricultural decision support systems.
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.
Czimer, D. · Kaluzsa, P. · Kövendi, J. · Li, K. L. · Kapusi, B. · Pomozi, V. · Fülöp, K. · Nagy, B. · Benedek, C. · Varadi, A. · Varga, M.
Quantifying skeletal mineralization phenotypes in larval fish is complicated by the natural curvature of the notochord and by sample-to-sample variability in orientation, staining and imaging. Consequently, many studies rely on summary measures such as vertebral counts or total stain intensity. Here we present SCAMP (Spinal Calcification & Mineralization Profiler), an open-source, GUI-based Python tool that computationally straightens the curved notochord of Alizarin Red S-stained fish larvae and generates standardized mineralization profiles along the spinal axis. This approach reduces positional and shape variability, allowing direct, quantitative comparison of calcification patterns within and between experimental cohorts, without requiring programming expertise. We validate SCAMP using a zebrafish model of Pseudoxanthoma elasticum (abcc6aelu15/elu15), recovering genotype-specific differences in the intensity, extent and spatial distribution of ectopic calcification. Using SCAMP, we further show that inorganic pyrophosphate (PPi) supplementation of the medium suppresses ectopic notochord calcification, alters the anterior-posterior distribution of mineralized regions in homozygous mutants, and promotes mineralization at physiological vertebral sites. We also show that methylene blue, a routine antifungal additive in fish medium, reduces baseline calcification, with the most pronounced effects observed in heterozygous controls. SCAMP is freely available and has the potential to be adapted to other fish species used in skeletal and mineralization research.
Early and accurate identification of plant diseases is essential for improving crop productivity and ensuring food security. Many existing deep learning-based plant disease classification methods rely solely on leaf images collected from a controlled environment, which limits their applicability in real-world agricultural conditions where symptoms may be visually unclear and influenced by environmental factors. To address these challenges, this study discusses AgriFusionNet, a context-aware multimodal deep learning framework that integrates leaf images, textual symptom descriptions, and environmental data for robust plant disease classification. The proposed architecture employs EfficientNet-B0 for visual feature extraction, BERT for semantic representation of symptom descriptions, and a lightweight multilayer perceptron for modeling environmental factors such as temperature, humidity, rainfall, and soil moisture. Features from all three modalities are fused into a unified representation to train the CNN model. The model is trained and tested upon the Context-Aware Multimodal Augmented PlantVillage dataset covering 38 plant diseases and healthy classes. Experimental results show that AgriFusionNet gives an overall accuracy of 98.94% on the dataset Context-Aware Multimodal Augmented PlantVillage, with competitive precision and recall and F1-score. The multimodal framework facilitates the co-learning of visual, semantic, and contextual environmental representations and the analyses of the confusion matrix and feature interactions give insights into cross-modal relationships. The proposed approach aims to explore context-aware multimodal representation learning for agricultural AI applications, with emphasis on integrating complementary visual, semantic, and contextual information.
Reproduction assets foundThe paper's data availability statement points to the Context-Aware Multimodal Augmented PlantVillage dataset (leaf images, symptom text, environmental data used for the phenotyping/classification analysis) deposited publicly on IEEE Dataport with a DOI matching an allowed URL.Dataset · publicPublicly available datasets were analyzed in this study. This data can be found here: Dataset. IEEE Dataport. https://dx.doi.org/10.21227/9jat-r836 [Accessed on August 2025].Open asset ↗IEEE Dataport · 10.21227/9jat-r836lines:1029-1047Plant phenotyping relevance matchCrossref · checked 5 Sept 2026
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.
Field / plotLeafObject detectionStress / disease detectionDisease symptoms / severity
Plant diseases are generally caused by pest, insects, pathogens and decrease the productivity to large scale if not controlled within time. Agriculturists are facing lose due to various crop diseases. It becomes tedious to the cultivators to monitor the crops regularly when the cultivated area is huge that is in acres. The proposed system provides the solution for regularly monitoring the cultivated area and provides the automated disease detection using remote sensing images. The proposed system intimates the agriculturist about the crop diseases to take further actions. The objective of the proposed system is to early detection of diseases as soon as it starts spreading on the outer layer of the leaves. The proposed system works in two phases: the first phase deals with training data sets. This includes, training both healthy and as well as diseased data sets. The second phase deals with monitoring the crop and identifying the disease using Canny?s edge detection algorithm.
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.
Cotton leaf diseases significantly affect crop productivity, fiber quality, and agricultural sustainability, making early and accurate disease diagnosis essential for effective crop management. Manual disease identification is time-consuming, subjective, and dependent on expert knowledge, highlighting the need for automated and intelligent diagnostic systems. This study proposes a CNN-based deep learning framework for automated multiclass classification of cotton leaf diseases using digital image analysis. The proposed framework is trained on a curated dataset comprising healthy and diseased cotton leaf images representing multiple disease categories. CNN architecture performs hierarchical feature extraction, automatically learning discriminative spatial representations from input images without manual feature engineering. The network is trained for 100 epochs using optimized hyperparameters to achieve robust feature learning while ensuring high generalization capability. Model performance is quantitatively evaluated using standard classification metrics, including accuracy, precision, recall, F1-score, and confusion matrix analysis, providing a comprehensive assessment of classification effectiveness. Experimental results demonstrate that the proposed framework achieves reliable disease recognition with high classification accuracy and minimal inter-class confusion, validating the effectiveness of CNN-based feature learning for agricultural image analysis. The automated system enables rapid and accurate disease diagnosis, supporting timely intervention, optimized crop protection strategies, and improved decision-making in precision agriculture. Future work will focus on improving model robustness through the integration of larger and more diverse datasets, advanced data augmentation techniques, transfer learning, and hybrid deep learning architectures. Furthermore, deployment of the proposed framework on mobile, edge, and web-based platforms, together with IoT technologies, can facilitate real-time field monitoring and disease surveillance. Overall, the proposed CNN framework provides a scalable, computationally efficient, and intelligent solution for automated cotton leaf disease classification, contributing to the advancement of AI-driven precision agriculture and sustainable crop management.
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.
Traditional methods for determining starch content in corn kernels are labor-intensive, destructive, and inefficient. To overcome these challenges, this work developed a rapid, non-destructive approach based on near-infrared hyperspectral imaging, applied to 58 rainfed corn varieties. A spectral preprocessing scheme combining wavelet transform, multiplicative scatter correction, and standard normal variate transformation was employed to enhance spectral quality. A two-stage wavelength selection framework was established using competitive adaptive reweighted sampling and sparrow search algorithm optimization. From the selected optimal wavelengths, four predictive models, namely partial least squares regression, artificial neural network (ANN), convolutional neural networks, and gradient boosting decision tree, were established, implemented, and systematically compared. The results identify 14 key wavelengths (1020.65-1647.71 nm) strongly correlated with starch content, with clear assignments to specific chemical bonds and good physical interpretability. Among these models, the ANN exhibited the best performance. The R 2 , RMSE, and RPD of the test set were 0.826, 0.759%, and 2.40, respectively, indicating favorable prediction accuracy and generalization ability. These key wavelengths provide a foundation for developing portable detection instruments. This work supports corn quality grading, breeding of high-starch varieties, and rapid raw material screening, thereby enhancing the quality and efficiency of the corn industry.
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.
Izabelle de Lima e Lima · Marta Laura de Souza Alexandre · Rodnei Rizzo · Ana Karla da Silva Oliveira · Carlos Augusto Alves Cardoso Silva · Peterson Ricardo Fiorio
Proper nitrogen (N) management is essential for increasing the productivity of sugarcane (Saccharum spp.) and reducing the economic and environmental impacts associated with excessive fertilizer use. This study compared the performance of two portable spectroradiometers, FieldSpec 3 and HandHeld 2, in estimating foliar nitrogen content based on hyperspectral data in the visible and near-infrared regions, obtained throughout the crop cycle. The experiment was conducted in Piracicaba, São Paulo, Brazil, under four N rates: 0, 60, 120, and 180 kg ha−1. Spectral measurements were taken at the foliar and canopy levels at eight evaluation times, accompanied by laboratory determination of N content. Partial Least Squares Regression (PLSR) and Random Forest (RF) models were fitted using the spectral data and days after cutting (DAC), included as a categorical factor and evaluated using 10-fold internal cross-validation, based on the metrics R2, RMSE, MAE, and Willmott’s refined agreement index (dr). The foliar data performed better with PLSR (R2 = 0.727; RMSE = 1.381 g kg−1; MAE = 1.109; dr = 0.917) than canopy data (R2 = 0.591; RMSE = 1.489 g kg−1; MAE = 1.157; dr = 0.866). PLSR also outperformed RF at both acquisition levels. The green (~550 nm) and red edge (~740 nm) regions were the most relevant for N estimation. Under the evaluated conditions, model performance was associated with the spectral acquisition level and conditions, the instrumental configuration, and the modeling strategy employed.
Doubled haploid (DH) technology significantly shortens the breeding cycle for developing homozygous inbred lines in maize ( Zea mays ). Manual sorting of haploids from a larger bulk of hybrid kernels in an induction cross is a major bottleneck in DH development. Automated systems based on near-infrared (NIR) reflectance spectroscopy can be valuable tools for rapid haploid sorting, provided that sorting accuracy is sufficient for incorporation into the DH process. In this study, we evaluated the accuracy of a custom-built single-kernel NIR (skNIR) sorter for classifying haploid kernels from 12 high-oil haploid induction populations generated from two sweet corn and two field corn donors and four high-oil haploid inducers (HOHIs). We evaluated several general classification models that can be applied without population-specific recalibration or prior genotyping, including models that classified haploids based solely on predicted oil content, as well as multivariate methods that used all wavelengths of the NIR spectra. The highest classification accuracy was obtained using a general multivariate support vector machine (SVM) model. When combined with the two best-performing HOHIs, the general SVM model accurately sorted induction populations from two of the three donor backgrounds crossed with these inducers. Two oil-based methods showed less accurate classification than the multivariate SVM model, due to overlapping oil content distributions across the two kernel classes. Overall, this study demonstrates effective skNIR-based sorting of haploid kernels from diverse induction populations using a single general model. The practical deployment of this instrument in maize breeding programs is discussed.
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.
In India, to increase the wheat yield production and support sustainable agricultural practices, timely and correct identification of wheat crop diseases such as yellow rust, brown rust, etc., is very much essential. Further, the majority of deep learning algorithms show high efficiency in plant disease detection and recognition, but these studies rely on handcrafted datasets captured under controlled laboratory conditions and accordingly will not perform well in the real world domain. Furthermore, there is a lack of real field wheat crop disease datasets collected in India, particularly from Haryana state, being the largest producer of wheat crops. Therefore, to address the issue of lack of region specific crop diseases dataset, which are region-specific, the wheat crop disease data set is curated, consisting of 2672 images of healthy and diseased leaves of wheat crops collected from the real field of CCS Haryana Agricultural University, Hisar. The dataset includes typically three classes of leaves of wheat crop, i.e., Yellow Rust (910), Brown Rust (922) and Healthy leaves (840) collected using both a DSLR camera and different smartphone cameras under natural lighting and field conditions. All images were validated and labeled by taking the expertise of wheat pathologists to ensure reliability on the dataset. The study also provides a comprehensive, systematic workflow for transforming raw data into a high-quality benchmark dataset for training using image preprocessing and augmentation techniques. Besides this, a comparative benchmarking analysis is performed under identical experimental conditions using three widely adopted deep learning architectures, which includes ResNet50, MobileNetV2 and EfficientNet-B0. The experimental results show that ResNet50 and EfficientNet-B0 achieve similar performance i.e. approximately 91% classification accuracy on real-field data. However, MobileNetV2 offers a lightweight alternative suitable for mobile and edge deployment. Furthermore, Grad-CAM based explainability analysis was performed to validate model predictions and highlight disease specific regions in wheat leaves. Therefore, this study contributes a practical region specific WheatDisease-HRY dataset and baseline benchmarking framework for developing robust AI based wheat disease diagnosis tools for real world agricultural applications in India.
The intelligent identification of tea diseases is crucial for ensuring tea quality and reducing economic losses in the tea industry. However, the deployment of deep learning models on edge devices remains challenging due to the conflict between detection accuracy and computational overhead. To address this, we propose CA-YOLOv8n, a lightweight object detection model tailored for tea disease diagnosis. Specifically, we introduce a Path-Decoupling strategy to streamline the network structure and integrate the Coordinate Attention (CA) mechanism to enhance the model's spatial awareness of subtle pathological features. Experimental results demonstrate that the proposed model achieves a mean Average Precision (mAP@50) of 98.89% while reducing the parameter count by 32.6% and FLOPs by 24.1% compared to the baseline YOLOv8n. The model was integrated into a diagnostic platform with an automated reporting interface, demonstrating that real-time tea disease identification is feasible on commodity CPU hardware in resource-constrained agricultural environments.
Reproduction assets foundThe paper's tea-leaf disease image dataset (9,591 images, YOLO format) is publicly deposited on figshare under CC BY 4.0, as stated in the Data Availability statement and dataset description. No author analysis code or trained model checkpoints are explicitly deposited.Dataset · publicAll data underlying the findings of this study are publicly available on figshare at https://doi.org/10.6084/m9.figshare.32253357 (CC BY 4.0).Open asset ↗figshare · 10.6084/m9.figshare.32253357lines:1-122Plant phenotyping relevance matchCrossref · checked 15 Sept 2026
Abstract Uneven occurrence of potato leaf diseases necessitates site-specific treatment rather than conventional uniform spraying. This study presents a vision-guided autonomous variable-rate spraying system integrating real-time deep learning-based disease detection, temporal disease severity estimation, and PWM-controlled precision spray actuation for targeted crop protection. A composite dataset comprising 1,561 field images containing 14,166 annotated leaf instances was developed under dense canopy conditions. Comparative evaluation of object detection architectures showed that YOLOv26 achieved the highest detection performance with a peak mAP@0.5 of 0.961, outperforming YOLOv8 (0.924) by 3.7% and YOLOv12 (0.938) by 2.3%. The lightweight YOLOv26n variant was selected for embedded deployment. When deployed on a Raspberry Pi 5, the optimized ONNX model achieved an inference speed of approximately 10 FPS, corresponding to a spatial sampling interval of 0.018 m at an operating speed of 0.18 m/s. A 2 s temporal sliding window generated a stable canopy-level disease severity index, which was mapped to PWM duty cycles for variable-rate pesticide application. Field experiments involving 598 spatial spray decision instances validated the proposed perception-driven spray control framework. Compared with conventional constant-rate spraying, the system reduced agrochemical consumption by 32.75% (from 458 to 308 L/ha). These results demonstrate the feasibility of integrating embedded deep learning and adaptive spray control into a field-deployable autonomous spraying platform for precision potato disease management.
Premise Accurate species identification is crucial for ecological restoration and can be especially challenging for understudied non-model species. Quercus garryana is the only native oak species in the Pacific Northwest and is an important component of the endangered oak savanna ecosystem. Quercus robur is an imported ornamental species from Europe and has been found to be mistakenly planted as Q. garryana in habitat restoration projects. Methods We measured leaf morphological traits sampled from herbarium collections in their native ranges using the digital morphometric tools MorphoLeaf and Tomato Analyzer. We then used Lasso logistic analysis to generate a predictive model and tested it on leaves from Portland, Oregon. To streamline this species detection process, we developed Garryanalyzer, an ImageJ plug-in that automatically measures leaf traits and outputs species predictions. Results Garryanalyzer demonstrated 95% accuracy in predicting the species identity of herbarium specimens of oaks. Garryanalyzer correctly identified all Q. robur individuals sampled in Portland but showed lower accuracy for Q. garryana . Discussion Many existing morphometric software are not open source, which makes them unable to be customized to specific study systems. Garryanalyzer is built upon the widely used open-source ImageJ platform. This study also demonstrates a viable workflow for developing similar tools for other ecologically important non-model plant species.
Reproduction assets foundThe paper's authors publicly released the Garryanalyzer ImageJ plug-in source code on GitHub, all original and modified leaf images used in the morphometric analyses on Zenodo, and the full leaf morphometric measurement dataset plus R Lasso analysis code in a second Zenodo repository. All are paper-specific, public,可直接Code · publicThe source code and installation instructions for Garryanalyzer can be accessed on GitHub at https://github.com/zxie8561/Garryanalyzer.Open asset ↗https://github.com/zxie8561/Garryanalyzer · zxie8561/Garryanalyzerhtml-lines:210-274Dataset · publicAll images used in the morphometric analyses, both original and modified, are available on Zenodo (https://doi.org/10.5281/zenodo.17462266).Open asset ↗https://doi.org/10.5281/zenodo.17462266 · 10.5281/zenodo.17462266html-lines:210-274Dataset · publicThe full dataset of leaf morphometric measurements of both GBIF and Portland samples, R code for Lasso analysis, and other miscellaneous files are available on a separate Zenodo repository (https://doi.org/10.5281/zenodo.17546152).Open asset ↗https://doi.org/10.5281/zenodo.17546152 · 10.5281/zenodo.17546152html-lines:210-274Plant phenotyping relevance matchOpenAlex · checked 11 Sept 2026
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.