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.
Site-specific spraying in chili pepper production requires organ-level segmentation of leaves, peppers, and flowers from handheld field images. However, RGB appearance becomes unreliable under organ overlap, occlusion, dust, and fruit specularity. Offline monocular depth from Depth Anything V2 provides an accessible structural prior without RGB-D sensing, but its reliability is spatially concentrated rather than uniform. To address this, we propose Depth-Routed Selective Attention (DRSA), an estimated-depth-guided segmentation network. DRSA predicts a single per-pixel routing field that, through one shared decision, jointly governs where depth-boundary cross-attention and residual depth fusion contribute, so geometric cues act near organ contours while RGB remains the default carrier. The routing field is calibrated online from depth-on and depth-suppressed predictions without trust-map annotation. We construct PepperField-EstDepth, a self-built dataset of 3, 940 handheld field images paired with estimated monocular depth, on which DRSA achieves \(90.20\%\) mIoU and \(84.48\%\) boundary mIoU, outperforming both RGB-only baselines and attention-based RGB-D fusion baselines; over the RGB segmentation reference, the gains are \(+1.98\) and \(+2.67\) percentage points, respectively. Under group cross-validation, DRSA reaches \(0.8919\pm 0.0031\) mIoU and \(0.8294\pm 0.0046\) boundary mIoU. For the spraying application, DRSA attains a target recall of 0.9814, a target precision of 0.9756, and an organ-level off-target activation of \(2.44\%\) . Visual-pipeline timing shows a segmentation-only latency of 43.0 ms under pre-generated estimated depth, rising to 219.6 ms for the RGB-to-mask visual pipeline when online Depth Anything V2-L depth generation is included. These results support DRSA as a pre-spray organ-level perception module.
Early and real-time detection of rice leaf diseases (RLD) poses a significant challenge for farmers, especially in rural regions with limited access to advanced technology. Conventional deep learning models often require substantial computational resources, rendering them impractical for deployment on mobile or edge devices commonly used in agricultural environments. Although models such as ResNet50, VGG16, and InceptionV3 can achieve high accuracy, they are computationally expensive and may be less suitable for real-time deployment in smart agricultural systems. Furthermore, many existing studies rely on limited datasets, which can restrict model generalizability under diverse real-world conditions. Although transfer learning can improve classification performance, developing lightweight models suitable for real-time IoT deployment remains challenging. To address these limitations, we curated a hybrid dataset of 7,092 rice leaf images by combining self-collected and Kaggle samples and proposed IoT-RiceMobileNet, a lightweight improved MobileNetV2-based model that can be effectively integrated into our developed IoT system. The proposed model outperformed transfer learning and deep learning baseline models, achieving 99.19% test accuracy and 99.20% precision while maintaining low computational complexity. We further confirmed model stability using stratified 5-fold and 10-fold cross-validation on both the constructed RLD and multi-source datasets, achieving mean accuracies of 98.04% and 98.13% on the constructed RLD dataset and 98.11% and 98.58% on the multi-source dataset, respectively. Moreover, our models compact size and 85.17 FPS inference speed support real-time deployment. Finally, the model was integrated into an IoT-enabled mobile, web, and cloud-based inference framework, demonstrating its practical potential for scalable rice disease detection in smart agriculture.
Reproduction assets foundThe paper's self-collected rice leaf image dataset and source code are publicly archived on Zenodo and GitHub, and the third-party Kaggle/Mendeley image datasets used to construct the RLD and multi-source datasets are publicly available. All are paper-specific, public, and actionable.Dataset · publicrice disease monitoring and decision support in precision agriculture.
Supporting information
S1 Appendix
Cross-validation results for the multisource dataset.
(PDF)
Data Availability
The self-collected rice leaf image dataset used in this study is publicly available through the archived IoT-RiceMobileNet repository on Zenodo: https://doi.org/10.5281/zenodo.21140529 . The public third-party datasets used in this study are available from their original sources: https://www.kaggle.com/datasets/dedeikhsandwisaputra/rice-leafs-disease-dataset/data
https://www.kaggle.com/datasets/anshulm257/rice-disease-dataset
https://data.mendeley.com/datasets/hx6f852hw4/2 Third-party images are not redistribOpen asset ↗Zenodo · 10.5281/zenodo.21140529lines:868-883Code · public://www.kaggle.com/datasets/anshulm257/rice-disease-dataset
https://data.mendeley.com/datasets/hx6f852hw4/2 Third-party images are not redistributed in the Zenodo archive and should be obtained from their original sources according to their respective licenses. The source code is publicly available through the GitHub repository: https://github.com/eimran/IoT-RiceMobileNet and is also archived on Zenodo at: https://doi.org/10.5281/zenodo.21140529 .
Funding Statement
The author(s) received no specific funding for this work.
References
1. Sokra I, Somaly S, Meta H, Sarun H, Molikoy C. Factors affecting rice production: A systematic review. J Agric Technol. 2026;2(1):19–46. doi: 10.6084/m9.figshaOpen asset ↗GitHub · eimran/IoT-RiceMobileNetlines:868-883Dataset · publicthe multisource dataset.
(PDF)
Data Availability
The self-collected rice leaf image dataset used in this study is publicly available through the archived IoT-RiceMobileNet repository on Zenodo: https://doi.org/10.5281/zenodo.21140529 . The public third-party datasets used in this study are available from their original sources: https://www.kaggle.com/datasets/dedeikhsandwisaputra/rice-leafs-disease-dataset/data
https://www.kaggle.com/datasets/anshulm257/rice-disease-dataset
https://data.mendeley.com/datasets/hx6f852hw4/2 Third-party images are not redistributed in the Zenodo archive and should be obtained from their original sources according to their respective licenses. The source code is Open asset ↗Kaggle · dedeikhsandwisaputra/rice-leafs-disease-datasetlines:868-883Dataset · publicataset used in this study is publicly available through the archived IoT-RiceMobileNet repository on Zenodo: https://doi.org/10.5281/zenodo.21140529 . The public third-party datasets used in this study are available from their original sources: https://www.kaggle.com/datasets/dedeikhsandwisaputra/rice-leafs-disease-dataset/data
https://www.kaggle.com/datasets/anshulm257/rice-disease-dataset
https://data.mendeley.com/datasets/hx6f852hw4/2 Third-party images are not redistributed in the Zenodo archive and should be obtained from their original sources according to their respective licenses. The source code is publicly available through the GitHub repository: https://github.com/eimran/IoT-RiceMOpen asset ↗Kaggle · anshulm257/rice-disease-datasetlines:868-883Dataset · publicived IoT-RiceMobileNet repository on Zenodo: https://doi.org/10.5281/zenodo.21140529 . The public third-party datasets used in this study are available from their original sources: https://www.kaggle.com/datasets/dedeikhsandwisaputra/rice-leafs-disease-dataset/data
https://www.kaggle.com/datasets/anshulm257/rice-disease-dataset
https://data.mendeley.com/datasets/hx6f852hw4/2 Third-party images are not redistributed in the Zenodo archive and should be obtained from their original sources according to their respective licenses. The source code is publicly available through the GitHub repository: https://github.com/eimran/IoT-RiceMobileNet and is also archived on Zenodo at: https://doi.org/10.5Open asset ↗hx6f852hw4/2lines:868-883Code / dataset availability confirmedOpenAlex · Crossref · checked 17 Sept 2026
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.
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.
Proline is an important biomarker in biological regulation and plant stress adaptation, increasing the demand for highly specific analytical detection methods. Herein, we report the use of DMAC as a chemodosimetric probe that exclusively identifies proline in the presence of 19 competing amino acids. The sensing mechanism relies on an irreversible chemical reaction between proline and DMAC, producing a distinct optical response with negligible background interference. This exhibits a highly sensitive limit of detection of 15.8 nM (1.82 ppb), signifying its importance in trace-level quantification. We thoroughly validated the underlying reaction mechanism using a combination of experimental and theoretical techniques. To demonstrate real-world applicability, the DMAC probe was integrated into a microfluidic device, resulting in a portable device designed for rapid, on-site analysis. Finally, the sensor's practical utility was demonstrated by successfully monitoring proline dynamics during plant growth.
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.
Introduction Accurate extraction of maize seedling phenotypes is essential for early growth assessment, yet field-grown seedlings exhibit complex backgrounds, non-rigid leaf deformation, self-occlusion, and adhered leaves that challenge three-dimensional point-cloud analysis. Methods Here, we developed a workflow integrating PMSPA-Net, an enhanced PointNet++ model, with geometric region decomposition for semantic and organ-level segmentation of field-grown maize seedlings. Multi-view RGB images of 65 complete plants were reconstructed with OpenMVS, and all data were partitioned at the whole-plant level before point-block sampling. Fifteen plants were retained as a fixed independent test set, while the remaining samples were used for training and validation across three predefined random seeds. Results Under a common protocol, PMSPA-Net outperformed PointNet, PointNet++, DGCNN, and RandLA-Net, achieving an mIoU of 0.8286 ± 0.0149, a macro-F1 of 0.8981 ± 0.0238, and a balanced accuracy of 0.9561 ± 0.0223. A four-configuration ablation study showed complementary gains from the multi-scale spatial pyramid attention module and IOA-based hyperparameter optimization. Geometric region decomposition also achieved more accurate and complete adhered-leaf separation than DBSCAN in three annotated cases. Point-cloud-derived leaf length and width agreed well with manual measurements. Discussion This workflow demonstrates the feasibility of integrated semantic and organ-level phenotyping under the evaluated field conditions; validation across additional sites, developmental stages, and acquisition conditions is required before broader generalization.
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.
Abstract Leafhopper feeding-damage severity is an important factor in the raw-material grading and final quality of Zijin Chan tea, but conventional assessment relies heavily on subjective visual judgment. This study developed a two-stage workflow combining machine-vision screening with offline near-infrared (NIR) spectroscopic reassessment. First, an improved small You Only Look Once version 8 (YOLOv8s) model localized tea-leaf targets and classified slight, moderate, and severe feeding damage. Physical samples corresponding to machine-vision no-result targets were then reassessed using an NIR model integrating standard normal variate preprocessing, Pearson correlation-based spectral-region selection, the successive projections algorithm, and a support vector machine. On 200 test images containing 4,000 annotated targets, the improved YOLOv8s model achieved precision of 88.6%, recall of 87.7%, and mean average precision at an intersection-over-union threshold of 0.5 of 89.3%, representing improvements of 12.2, 9.3, and 8.7 percentage points over the baseline, respectively. Developed from 240 independent physical samples, the NIR model achieved macro-averaged recall of 94.03% on the cross-validated training set and 95.12% on the spectral model-selection set. When the fixed model was applied to physical samples corresponding to 492 machine-vision no-result targets, 473 targets were classified correctly, yielding a reassessment accuracy of 96.14%. Across all targets, standalone machine vision achieved an end-to-end accuracy of 77.7%, whereas the two-stage workflow achieved 89.5%. These results support the feasibility of combining visual and spectral information for laboratory-based grading. Validation using in situ spectra and samples from multiple batches and seasons remains necessary before online deployment.
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.
Diseases that affect plants considerably lower the production of crops and pose a significant threat to global food security, particularly in high-value crops like bell pepper (Capsicum annuum L.). The timely and sensitive identification of the disease is essential in reducing losses related to yield and paving the way to sustainable management of crops. Although automated plant disease diagnosis has been improved by artificial intelligence (AI)-driven image analysis, single-model methods are not always robust or generalisable. The study proposes a hybrid convolutional neural network-random forest (CNN-RF) framework for the early detection of bell pepper leaf diseases. Deep visual features of leaf images were extracted and subsequently used with a pretrained ResNet50 model and then used to classify the images with the help of a RF ensemble algorithm. Experiments were conducted on a publicly available PlantVillage bell pepper dataset comprising 495 leaf images (199 bacterial spot and 296 healthy samples). The classification accuracy of the proposed hybrid model was 97.98 % with high class-wise precision, recall and F1-scores. Receiver operating characteristic (ROC) analysis produced an area under the curve (AUC) of 0.98, indicating excellent discriminative performance. The dataset was evaluated using a stratified 80 : 20 train-test split. The results show that deep feature extraction combined with an ensemble learning method can improve classification accuracy and offer a scalable solution to assist in precisionagriculture and timely disease control. However, the framework was evaluated only on the PlantVillage dataset and has not yet been validated under real-field agricultural conditions. The framework demonstrates potential for extension to multi-disease classification tasks, although further validation on field-acquired datasets is required.
Double fertilization in angiosperms is completed by the fusion of two sperm cells with the egg and central cells. While sperm membrane proteins such as GEX2, DMP8/9, and GCS1/HAP2 have been identified as key regulators of gamete attachment and fusion, their specific ultrastructural roles within the embryo sac have remained unclear owing to the technical limitations of live-cell imaging based on fluorescent markers. Here, we optimized high-pressure freezing and freeze-substitution protocols for Arabidopsis pistils and performed three-dimensional reconstructions using scanning transmission electron microscopy (STEM) and array tomography. We identified an electron-dense, partially thickened structure at the apical region of the egg cell before pollination. This structure disappeared following wild-type fertilization, suggesting it serves as a specialized "docking site" to accommodate pollen tube contents and facilitate subsequent gamete interactions. Comparative analysis of fertilization mutants revealed distinct functional stages in the gamete fusion process. In gex2 mutants, sperm cells failed to establish robust connections when positioned at the lateral side of the egg cell, confirming the requirement of GEX2 for initial attachment. In contrast, dmp8 dmp9 double mutant sperm cells facing each other were observed, exhibiting a more distant position from the egg surface, suggesting that DMP8/9 is involved in maintaining stable membrane adhesion before fusion. Furthermore, although gcs1 sperm cells maintained intimate contact and displayed membrane thinning, they failed to undergo final fusion. Our findings demonstrate that STEM and array tomography are powerful tools for visualizing gamete dynamics, providing an integrated model of the transition from stable adhesion to membrane fusion.
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.
PotatoField / plotPhysiological trait estimationGrowth / time-series analysisBiomass / plant weightGrowth / development / phenologyWater status / transpiration
Real-time measurement of belowground tuber growth has not been conducted in field crops. Here, a strain-gauge sensor was used to monitor the growth of a single potato tuber and estimate its daily water loss. Over the 11 days preceding harvest, tuber thickness increased by 0.80 mm, corresponding to a daily gain of 1.4 g, or a 4.2% increase. The greatest diurnal fluctuation was 0.439 mm, corresponding to a transpirational water loss of 9.1 mL, or 2.9% of the tuber’s water content. Daily estimated transpiration showed a positive correlation with estimated vapor pressure deficit. This sensor may enable more precise input control and higher temporal resolution than current methods for below-ground crops, supporting improved crop management, yield prediction, and harvest decisions.
This study aimed to evaluate the growth responses of kenaf (Hibiscus cannabinus L.) under nitrogen and compost treatments and to develop a UAV-based plant height estimation model. Ground-measured plant height differences were not significant at harvest (110 days after sowing, DAS), highlighting the need for multitemporal monitoring. Multispectral drone imagery was acquired at five growth stages (20–110 DAS). Object-based image segmentation was applied to extract pure vegetation areas, and digital surface model differencing (ΔDSMt) was used to reduce micro-topographic effects. A UAV-based multiple linear regression (UAV-MLR) model was developed using ΔDSMt, NDVI, GNDVI, and NGRDI to integrate complementary structural and spectral information. Evaluated on the calibration dataset, the UAV-MLR model demonstrated high fitting performance (adjusted R2 = 0.9858, RMSE = 14.95 cm, MAE = 11.31 cm), outperforming the ground-based simple linear regression (G-SLR) model based on stem diameter (adjusted R2 = 0.9615, RMSE = 39.68 cm, MAE = 29.16 cm). By integrating structural and spectral information, the proposed approach reduced RMSE by 62.3%, offering a highly accurate, non-destructive tool for crop monitoring and precision agriculture.
Abstract 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.
Plant cells grow and differentiate in an ever-changing environment characterized by transient signals and stimuli. The ability of plant cells to perceive, integrate, and dynamically respond to these stimuli underpins a plant adaptation and survival. Among the plethora of complex stimuli plant cells are exposed to, several stimuli have a mechanical component, for example, wind, touch, contact with insects, penetration of pathogens, and even intrinsic mechanical stresses arising during tissue growth. Despite the presence of a load-bearing cell wall that separates cells from the environment and fixes their location within a tissue, plant cells are responsive to mechanical stimuli. However, several questions remain unanswered around how mechanical stimuli are perceived and translated into cellular responses. Here we establish a system enabling application of quantifiable localized mechanical stress while simultaneously capturing cellular responses with high spatio-temporal resolution using confocal imaging. We show that this system enables estimation of locally applied pressure and provides access to the temporal and spatial details of subcellular events in intact living tissues upon touch. We propose that observing these subcellular events at high spatiotemporal resolution and linking their dynamics to the intensity of mechanical stimuli may uncover the molecular mechanisms underlying plant cell responses to touch.
Reproduction assets foundThe paper's quantitative analysis code is publicly available: the authors state that R scripts for load-cell force readout analysis and Fiji/R scripts for fluorescence signal profiling are available at the Microindentation-toolbox GitHub repository, and the LabView load-cell readout software (Flamos) used for the paperCode · publicpush on
134
the mobile plate of the load cell. The needle was then lifted and lowered again after a
135
few seconds interval to create repeated touches on the load cell mobile plate. Force
136
readouts from the software were then analyzed using an R (R Core Team, 2025) script
137
(publicly available at the Github repository
138
https://github.com/AnnalisaBe/Microindentation-toolbox.git). For pressure estimates,
139
the plant-needle interface was approximated to the curved surface of a half sphere.
140
The pressure applied on the tissue upon needle movement was then estimated as the
141
force exerted by the needle movement (as calibrated by the load cell) divided by the
142
curved surface of Open asset ↗AnnalisaBe/Microindentation-toolboxpdf-raw-page:4 lines:1-88Plant phenotyping relevance matchbioRxiv · Europe PMC · checked 15 Sept 2026
Leaves of terrestrial plants possess heterogeneous anatomical structures composed of multiple cell layers. Although biochemical differences among leaf tissues have been inferred from protein analyses and anatomical studies, direct comparisons of dynamic photochemical responses among tissues while preserving their spatial context remain challenging. Leaves experience intrinsically heterogeneous environments because incident light enters primarily from above and propagates through complex internal leaf structures. Therefore, analyzing photosynthetic activity at the tissue and cellular levels is essential for understanding how photosynthesis operates within structurally heterogeneous leaves. Here, we combined live leaf-section imaging with microscopic Imaging-PAM chlorophyll fluorescence measurements to analyze photochemical responses at the tissue level in living leaf sections. In dorsiventral dicot leaves, palisade tissues exhibited a higher effective PSII quantum yield [Y(II)] and more rapid induction of regulated energy dissipation [Y(NPQ)] than spongy tissues, indicating higher photosynthetic capacity and photoprotective activity. In contrast, rice leaves, which lack palisade-spongy differentiation, showed uniform photochemical responses along the adaxial-abaxial axis. In the C4 plant finger millet, mesophyll and bundle sheath cells displayed distinct photochemical responses consistent with their functional differentiation in C4 photosynthesis. These results demonstrate that photosynthetic responses are spatially organized according to leaf anatomical structure. Although section-based measurements do not reproduce the native optical and gas environments of intact leaves, they enable the comparison of intrinsic tissue-specific photochemical properties under approximately equivalent illumination. The present approach enables tissue-resolved chlorophyll fluorescence analysis in living leaves, providing a new framework for investigating how leaf architecture shapes the spatial organization of photosynthetic activity.
This study presents a systematic quantitative multi-elemental investigation of four major plant organs (roots, stems, leaves, and flowers) of Nerium oleander using calibration-based Laser-Induced Breakdown Spectroscopy (LIBS), with the results validated by Atomic Absorption Spectroscopy (AAS). Plasma characterization was carried out using Boltzmann plot and Stark broadening analyses, while negligible self-absorption observed through the Hα emission line confirmed optically thin plasma conditions and reliable quantitative measurements. A total of nine elements were detected, including Fe, Zn, Mn, Ca, Mg, K, Na, Cu, and Ni, each exhibiting different concentration levels across the analyzed tissues. Compositional analysis using standard calibration curve-based LIBS demonstrated that elemental concentrations were non-uniform, showing marked variations between the different plant tissues. Among the detected elements, calcium emerged as the most prevalent across all tissues. The highest calcium concentration was observed in leaves (16,385 mg L-1), followed by roots (12,092 mg L-1) and flowers (11,185 mg L-1). Root tissues exhibited elevated concentrations of Fe and Mn, reaching 1918 and 513 mg L-1, respectively. In contrast, flowers showed the highest Mn concentration (1888 mg L-1), while leaves were enriched in Mg (4445 mg L-1) and K (3700 mg L-1). The highest Na concentration was observed in stems (8025 mg L-1). Trace metals, including Cu, Zn, and Ni, were detected at comparatively low concentrations, while Pb remained undetected in all samples. The strong agreement between LIBS and AAS measurements confirms the reliability of the proposed methodology and demonstrates the potential of LIBS as a rapid and non-destructive tool for elemental assessment of medicinal plants.
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:1-52Code / dataset availability confirmedEurope PMC · checked 15 Sept 2026
Kawai T, Teramoto S, Ma X, Fukushima D, Hmwe KK, Kimani SM, Tokida T, Uga Y.
RiceMultimodalX-ray / CTRootMorphology / geometry measurementPhysiological trait estimationGrowth / time-series analysisGrowth / development / phenologyRoot system architecture
Rhizosphere oxidation is a key adaptive mechanism in reductive soil environments, in which oxygen released from roots alters rhizosphere redox conditions and regulates biogeochemical processes. Rice plants possess an internal oxygen transport system, and radial oxygen loss (ROL) from roots is closely associated with root development. However, the spatial patterns of ROL in soil and their relationships with root traits remain poorly characterized. In this study, we developed a multimodal imaging system that integrates planar oxygen optodes with X-ray computed tomography to simultaneously visualize rhizosphere oxidation and root development in rice. Daily time-course tracking of individual crown roots revealed dynamic changes in the spatial distribution and magnitude of rhizosphere oxygen in relation to root elongation and aging. Root thickness was positively correlated with dissolved oxygen levels near root tips. Genotypic comparisons further identified a cultivar with reduced rhizosphere oxidation despite possessing thicker roots among the tested genotypes, thereby indicating the involvement of additional physiological processes. Overall, these findings demonstrate that rhizosphere oxidation is regulated by root growth stage and thickness and dynamically modulated during root development.
Reproduction assets foundThe paper's Data availability statement explicitly deposits the authors' RG2DO-Root analysis program together with sample optode and CT images (the paper's phenotyping inputs) in a public GitHub repository, matching the allowed URL.Code · publicing
8
This work was supported by project JPNP18016, commissioned by the New Energy and
9
Industrial Technology Development Organization (NEDO), JST CREST (JPMJCR17O1),
10
and JST ALCA-Next (JPMJAN23D3).
11
12
Data availability
13
The source code and sample data (optode and CT images) are available from the
14
GitHub repository (https://github.com/tsubasa-kawai28/RG2DO-Root).15
16
References
17
Aguilar EA et al. 2003. Oxygen distribution and movement, respiration and nutrient
18
loading in banana roots (Musa spp. L.) subjected to aerated and oxygen-depleted
19
environments. Plant Soil. 253:91–102. https://doi.org/10.1023/A:1024598319404.20
Armstrong W, Wright EJ. 1975. Radial oxygen loss fromOpen asset ↗https://github.com/tsubasa-kawai28/RG2DO-Root · RG2DO-Rootpdf-raw-page:19 lines:1-82Plant 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.
Volatile organic compounds (VOCs) released by plant leaves play key roles in stress signaling, plant-atmosphere interactions, and plant defense. Among these, wound-induced VOCs (wVOCs) are emitted within seconds of mechanical damage, herbivory, or environmental disturbance. Their emission dynamics depend strongly on the timing, severity, and method of tissue disruption. Yet, accurate quantification remains challenging due to mechanical artifacts, variable exposure conditions, and delays between injury and measurement. This study presents a standardized within-chamber leaf excision protocol for real-time monitoring of wVOCs and gas exchange. A surgical-grade cutter was integrated into a portable gas-exchange chamber to enable clean, controlled cuts within a sealed chamber under stable light, humidity, CO2, and temperature conditions. A proton-transfer-reaction time-of-flight mass spectrometer (PTR-TOF-MS) continuously measured volatile emissions at the chamber outlet, minimizing delay and signal distortion. This setup resolves emission onset, peak timing, maximum rise rate, and total release with high temporal fidelity. Application of the method to Quercus rubra, Acer platanoides, and Gossypium hirsutum demonstrated its ability to resolve distinct wound-induced emission patterns across contrasting leaf types. By eliminating delays associated with conventional sampling, this method resolves the full kinetic trajectory of wound-induced emissions and overcomes major limitations of previous approaches. It provides a robust framework for studying rapid stress responses in plant physiology, ecological biochemistry, and plant-atmosphere interactions.
Transmission electron microscopy (TEM) is widely used to examine plant cellular ultrastructure, but sample preparation remains challenging because polysaccharide-rich cell walls, large vacuoles and complex membrane systems compromise staining and structural preservation. Here, we developed a plant-adapted TEM preparation method based on a modified osmium-thiocarbohydrazide-osmium (OTO) staining workflow combined with optimized washing, dehydration and resin infiltration. Using tomato roots and leaves, we show that the optimized method produces cleaner backgrounds, better-defined cellular boundaries and improved preservation of cellular morphology compared with conventional preparation. Fine membrane-associated structures, including mitochondrial cristae, chloroplast grana and stroma lamellae, nuclear membranes and other endomembrane structures, were more clearly resolved, while staining-related precipitates and diffuse background artifacts were reduced. This method provides a practical approach for high-contrast TEM imaging of plant tissues and ultrastructural analysis of plant cells and organelles.
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 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 matchOpenAlex · checked 15 Sept 2026
Quantitative measurements of insect flight behaviour are essential for understanding the biomechanics, control, and ecology of flight, yet obtaining such measurements under free-flight conditions remains challenging. Small body size, rapid wing motion, visual symmetry, and frequent occlusions complicate three-dimensional pose estimation, often requiring restrictive experimental setups or substantial manual annotation. We present FLiTrak3D, an open-source Python package for estimating insect flight kinematics from multi-view videography. It combines machine-learning-based markerless tracking with biomechanical modelling to reconstruct and parametrise insect body and wing motion. The workflow integrates image preprocessing, including dynamic image cropping and enhancement, two-dimensional bodypart localisation, three-dimensional reconstruction, and optimisation-based skeletal fitting. A key innovation is the use of spatio-temporal encoding across synchronised camera views and adjacent frames to improve neural-network awareness of spatial and temporal context during bodypart localisation. Multi-view image stitching allows the network to use cross-view spatial relationships, reducing left-right bodypart misidentifications, while temporal encoding stacks consecutive greyscale frames into RGB images provide short-term motion information. A species-specific skeleton is then fitted to reconstructed keypoints to enforce kinematic constraints, and estimate body and wing orientations. We demonstrate the approach using free-flying Aedes aegypti mosquitoes, achieving bodypart localisation errors close to human labelling, and realistic flight kinematics.
Zea mays (maize) is a critical global crop, generating more than 1 billion metric tons of grain annually. Plant morphology is tightly associated with final yield and is determined by the process of development. Accurate assessment and recording of mature plant morphology is essential for linking genotype to phenotype and for assessing environmental influences on plant development and yield. In maize, imaging of fully emerged aerial organs provides an accessible and informative developmental end point for comparative phenotypic analyses. Here, we provide a protocol for capturing high-quality images of mature maize plants and detached leaf blades, suitable for publication and downstream quantitative trait extraction. We emphasize standardized imaging conditions, including high-contrast backgrounds, diffuse lighting, fixed camera positioning, and consistent scale calibration, to minimize technical variation and to facilitate automated image processing. Guidance is provided for both laboratory- and field-based imaging setups and uses widely available equipment. Images generated using this protocol can be analyzed using platforms such as FIJI to extract morphometric traits, including length, width, and outline-based shape descriptors. By prioritizing standardization at the point of acquisition, this protocol supports reproducible, scalable phenotyping of mature maize morphology for developmental studies in maize.
In plants, final organ shape and position influence function and whole plant productivity. In Zea mays (maize), leaf, ear, and tassel morphology are all key agronomic features influencing yield. These aerial structures originate from stem cell populations within meristems. Understanding how meristems are regulated and how stem cells differentiate over developmental time to form mature leaves, ears, and tassels is important for identifying novel routes for maize improvement. High-resolution imaging of meristem morphology is a key component of developmental studies. Scanning electron microscopy (SEM) is widely used to visualize meristem topology; however, conventional SEM preparation requires fixation, dehydration, drying, and conductive coating. These steps are time-consuming, technically demanding, and can introduce damage. Here, we present a rapid protocol for imaging freshly dissected meristems from maize using a benchtop scanning electron microscope. Fresh tissues are mounted directly onto conductive adhesive and imaged under high vacuum at low accelerating voltage, enabling high-quality micrographs to be captured within minutes of dissection. Although dehydration and tissue collapse can occur under vacuum, rapid sample handling and image acquisition minimize these effects and produce images comparable to those obtained from conventionally prepared specimens. This streamlined workflow substantially reduces preparation time, lowers material requirements, and facilitates medium-throughput morphological analysis. The method provides a practical and reproducible approach for rapid developmental phenotyping and is readily adaptable to other plant species and other tissues.
The final morphology of a plant is determined by much earlier events that occur during development. Meristems are critical for the development of the aerial parts of plants, such as leaves and stems. The morphology of these meristems, which can vary significantly across genetic backgrounds, correlates with phenotypic traits in Zea mays (maize) that influence yield, such as leaf initiation rate and kernel row number. Traditional methods for morphological analysis of meristems, including histological sectioning and fine dissection, often require considerable technical expertise. To address these challenges, we present a detailed protocol employing methyl-salicylate for tissue clearing, combined with Differential Interference Contrast (DIC) microscopy to visualize maize shoot apical meristems. This method enables the intact visualization of vegetative meristems without the need for extensive dissection, thereby preserving the natural morphology of the meristem. Using this approach, the meristem is coarsely dissected, fixed, optionally stained, and cleared to enable visualization through leaf layers. By improving the accessibility and quality of meristem imaging, this protocol facilitates accurate quantification of meristem morphology, critical for genetic and developmental studies. Optimized primarily for maize, the outlined techniques can also be adapted for other plant species, providing a versatile tool in the field of plant developmental biology.
Advanced crop monitoring inside greenhouses is becoming one of the primary objectives of research centers. High-performance sensors, such as LiDAR or stereo cameras, have traditionally been employed for this purpose, though these often have a high cost. This work proposes a Visual-SLAM system using a monocular camera, which is significantly more cost-effective and specifically tailored for agricultural applications, such as mapping tomato crops in a greenhouse. Tests were carried out on a real tomato bunch, located in the Agroconnect experimental greenhouse. A ROS 2 Humble node was developed to run on the robot in order to capture images of these crops, which were then stored for offline processing. To generate a 3D mapped model for the crop in the greenhouse, the GLOMAP mapper, based on Structure-From-Motion, was integrated with the Hierarchical Localization toolbox. This initial mapping is a foundation for future, more advanced algorithms to analyze growth patterns, and optimize agricultural management. The system leverages a hierarchical localization paradigm based on a coarse-to-fine strategy: it first performs global retrieval to generate location hypotheses, then combines local features within the identified candidate regions. The results show a correct identification of the tomato cluster, correctly characterising the tomato that is occluded and inaccessible by classical vision technologies. The reconstructed 3D model was further validated against manual ground-truth measurements of fruit size, centroid position, and orientation, confirming the geometric accuracy of the proposed low-cost monocular pipeline.
Abstract Purpose Agrivoltaic vineyards show strong spatio-temporal variability in canopy shading, but field methods to quantify panel-induced shading at canopy scale remain limited. Shading is a key factor because it affects plant physiological and morphological traits, with potential consequences for yield and production quality. This study developed a near-surface time-lapse RGB imaging approach to derive temporally explicit shading metrics in an agrivoltaic vineyard of Vitis vinifera cv. Falanghina in Southern Italy. Methods Two representative vine positions beneath the photovoltaic structure were monitored: Agrivoltaic Shade (AVS), with greater exposure to panel-induced shading, and Agrivoltaic Light (AVL), with lower exposure. Image-based canopy shading percentage was calculated through a dedicated processing workflow and integrated with radiometric and physiological measurements, including continuous photosynthetically active radiation (PAR), canopy-level spectral photon flux measurements, photosynthetic photon flux density (PPFD), band-specific photon flux densities, red:far-red ratio (R:FR), stomatal conductance (gₛ), and leaf temperature. PAR measurements beneath the panels were compared with a full-sun control area. Results AVS showed significantly higher shading than AVL (76.14% vs 39.45%, p Conclusion The proposed workflow offers a low-cost, non-destructive tool to quantify shading dynamics and support site-specific assessment of crop microenvironments in agrivoltaic systems. The approach provides crop-relevant information for precision monitoring and management of spatially heterogeneous light conditions across different crop species. Impact The data provided in this manuscript enable the quantification of in-season photovoltaic-induced canopy shading dynamics in an agrivoltaic vineyard using proximal RGB time-lapse imaging and crop-level radiometric measurements. These metrics reflect the spatial and temporal variability of light availability within the vineyard and support site-specific assessment of crop microenvironments and precision management of agrivoltaic systems.
Polyploidy can be a critical factor for explaining plant trait variation, niche diversification, or speciation. However, inferring ploidy from silica-dried or historical samples using chromosome counts or flow cytometry is not possible, and scaling up ploidy estimation to population-level fresh contemporary samples can be challenging as well. Thus, we present a new method for estimating ploidy levels directly from sequencing data using machine learning; the Polyploid Population Genomics Tool Kit (PPGTK). The machine-learning approach is advantageous as it relaxes the assumptions of previous probabilistic methods and provides per-sample probabilities, allowing investigators to evaluate uncertainty in their system of interest.. We demonstrate performance and accuracy of the method on simulated and empirical data. Simulations showed above 99% accuracy, even for low coverage data, as long reads were mappable to the reference genome. For empirical analyses, we used target enrichment data from blueberry wild relatives (Vaccinium sect. Cyanococcus) and whole-genome data from sweetpotato wild relatives (Ipomoea ser. Batatas). Ploidy was recovered with 99% accuracy across 70 Vaccinium individuals and 97% across 82 Ipomoea individuals. Analysis of many individuals is fast and requires only a multisample VCF, which is presumably generated for the research anyway, and some samples of known ploidy for training the classifier. The approach implemented in PPGTK is promising for collections-based research as well, enabling ploidy classification of historical specimens based on present-day observations. The method is implemented in a new Python package as a single command that can run on a conventional laptop.
Reproduction assets foundThe paper's ploidy-classification method is implemented in the authors' public Python package PPGTK, with a specific release (v0.1.0-alpha) used for the manuscript's analyses. The empirical VCF/metadata datasets are promised on Dryad only 'upon acceptance' and thus are not yet actionable.Code · public11
VCFs and metadata needed to reproduce Vaccinium sect. Cyanococcus and Ipomoea ser.
372
Batatas analyses with PPGTK will be made available via Dryad upon acceptance. PPGTK is
373
available on GitHub, and release v0.1.0-alpha was the version used for analyses in this
374
manuscript (https://github.com/tileylab/PPGTK/releases/tag/v0.1.0-alpha). PPGTK currently has
375
other functions for calculating population genetic summary statistics, but the classify-ploidy
376
function implements the machine-learning method described in the manuscript.
377
378
.
CC-BY 4.0 International license
is made available under a
preprint (which was not certified by peer review) is the auOpen asset ↗tileylab/PPGTK · v0.1.0-alphapdf-raw-page:11 lines:1-24Plant phenotyping relevance matchEurope PMC · checked 15 Sept 2026
Balem JM, Tan C, Dias NCF, Arnold M, Tran S, Severns PM, Teixeira PJPL, Li C, Yang L.
ArabidopsisThermalLeafTissueGrowth / time-series analysisStress response / tolerancePlant / canopy temperature
Repairing damaged tissues is essential for the survival of all organisms. In plants, tissue injury rapidly triggers defense and repair programs. However, the molecular mechanisms linking early injury cues to the later stage of wound repair remain unclear. Here, we show that wounding of Arabidopsis leaves induces localized low temperature at the injury site, likely caused by evaporative cooling, which is accompanied by an activation of cold-responsive genes. Using thermal imaging combined with computer vision and deep learning, we developed a workflow to monitor the dynamics of wound healing in a quantitative, non-invasive, and real-time manner. Mechanistically, we show that C-repeat Binding Factor (CBF) transcription factors are required for the activation of the injury-associated cold response and downstream salicylic acid (SA) signaling. Our findings suggest that the CBF-SA pathway acts coordinately to promote lignin and callose deposition, thereby facilitating wound repair. Together, these findings reveal a link between a wound-induced biophysical cue and the tissue repair program.
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.
Kelsey M. Reed · Abigail A. Masri · Andrew J. Hanrahan · Emery L. Ng · Mao Li · Bastiaan O. R. Bargmann
ArabidopsisLaboratory / benchtopMicroscopyCell / cellular structureGrowth / time-series analysisTrackingVisualization / data managementGrowth / development / phenology
Plant developmental biology lacks cell-based experimental systems comparable to the organoids and live-imaging platforms that have transformed mechanistic discovery in animal research. To address this gap, we present a robust, trackable protoplast regeneration platform in Arabidopsis thaliana that enables high-resolution, time-resolved analysis from the single-cell stage through microcolony formation and early regenerative development. Protoplasts are embedded in thin alginate matrices containing fluorescent fiducial beads, maintaining physical separation and allowing repeated return to the same cells over days to weeks. This design supports long-term imaging using epifluorescence, confocal, and lattice light sheet microscopy, enabling visualization of cell-cycle re-entry, asymmetric division, organelle dynamics, dedifferentiation, redifferentiation, and regenerative competence. Fluorescent reporters for nuclei, membranes, microtubules, Golgi, and hormone signaling further permit observation of subcellular organization and signaling heterogeneity during early reprogramming. Together, this platform provides an accessible, scalable system for studying plant cellular plasticity at single-cell resolution and offers a foundation for plant organoid-like models. By enabling high-resolution study of regeneration from single cells, this method expands the experimental toolkit and supports efforts to overcome species- and genotype-dependent barriers to transformation and plant biotechnology.
Abstract Accurate estimation of maize leaf nitrogen content is important for improving nitrogen-use efficiency and supporting precision crop management. However, leaf-level hyperspectral modeling is challenged by high spectral redundancy and heterogeneous spectral responses among local leaf regions. This study proposes a Hyperspectral–Region Aggregation Network (HSRAN) for maize leaf nitrogen content estimation from region-level hyperspectral spectra. HSRAN consists of a Spectral Adaptive Recalibration Encoder (SARE) and a Context-Aware Gated Aggregation Module (CAGM). SARE performs band-wise residual recalibration and extracts regional spectral representations, whereas CAGM models contextual dependencies among regional features and performs gated attention-based aggregation for leaf-level prediction. Field experiments were conducted in 2024 and 2025 at the jointing, silking, and maturity stages. HSRAN was evaluated against PLSR, RF, XGBoost, SVR, 1D-CNN, MLP, and Transformer1D models. Across the stage-specific and pooled datasets, HSRAN achieved the highest R² and the lowest RMSE while maintaining competitive MAE values. On the pooled full-growth-period dataset, HSRAN achieved an R² of 0.84, an RMSE of 3.63 g kg⁻¹, and an MAE of 2.59 g kg⁻¹. At the jointing, silking, and maturity stages, the corresponding R² values were 0.56, 0.76, and 0.72, respectively. Ablation experiments indicated that integrating SARE and CAGM improved R² from 0.80 to 0.84. To interpret regional contributions, the learned attention weights were mapped back to the original leaf coordinates recorded during regional sampling. Regions near leaf veins, tips, and margins often received relatively higher attention weights, suggesting that their local spectra provided informative cues for model prediction. These findings indicate that spectral–regional joint modeling can improve leaf-level hyperspectral estimation of maize nitrogen content. HSRAN provides a practical framework for non-destructive nitrogen assessment in maize.
Doolan O, Peirats-Llobet M, Lewsey MG, Aberdein N, Bricklebank N.
BarleySeed / grain
Quantitative imaging of plant tissues by laser ablation inductively coupled plasma mass spectrometry (LA-ICP-MS) is hindered by the lack of matrix-matched calibration standards. Established approaches, such as gelatine or homogenised tissue blocks, do not replicate plant matrices accurately. Here, we introduce a nano-dispenser-based calibration strategy that deposits nanolitre volumes of elemental standards directly onto paraffin-embedded grain sections, exploiting the low endogenous metal content of the endosperm to generate in situ calibration curves. Calibration performance for Mg, Mn, Cu, Zn, and Mo was assessed using LA-ICP-MS imaging and Iolite 4 data processing. The method demonstrated excellent linearity (R 2 > 0.98), reproducibility across multiple grains, and sub-ppm limits of detection. Comparative analysis with in-house homogenised blocks and NIST wheat reference material confirmed superior accuracy and reproducibility of the nano-dispenser approach. As a proof-of-concept, we have applied the method for the quantitative imaging of metals in a whole barley grain section, and the results show excellent agreement with published data obtained by conventional liquid-mode ICP MS. The technique reported here provides a robust, scalable, solution for quantitative metallomics studies of plant tissues, enabling improved assessment of nutrient distribution and supporting the development of standardised protocols for LA-ICP-MS imaging.
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 · checked 15 Sept 2026
Volatile aroma compounds (VACs) are key semiochemicals induced and released by plants under biotic and abiotic stresses. Therefore, it is of great significance for detecting VACs to evaluate the growth environment information and physiological markers of plants. However, real-time and on-site monitoring of VACs at the current stage remains a challenge due to their complex species and trace concentrations. Here, we report an ultrasensitive VAC sensor based on Ru-nanocluster-anchored ZnO nanodisks, for which the corresponding limit of detection for VACs can reach as low as 10 ppb. By constructing temperature-modulated 4 × 4 VAC sensor arrays and integrating artificial intelligence algorithms, we demonstrate simultaneous identification of 10 VAC species and concentrations with 98% accuracy and high fidelity. Furthermore, a portable biomimetic olfactory e-nose based on temperature-modulated sensor arrays was developed to realize the on-site detection of VACs released from tea plants induced by Ectropis obliqua. This work not only provides a design strategy for ultrasensitive VAC sensors but also establishes a scalable platform for real-time plant health monitoring and early warning of biotic and abiotic stresses.
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:177-215Code / dataset availability confirmedCrossref · checked 15 Sept 2026
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.
Accurate litchi counting from whole-tree images is essential for yield estimation, orchard management, and plant phenotyping, but remains challenging in real orchards because fruits occur in dense, heavily occluded clusters and vary markedly in scale, illumination, and appearance across ripening stages, particularly when green fruits resemble surrounding foliage. Existing methods have shown promise, but their robustness in complex orchard environments remains limited. To address these challenges, we propose FG-LCNet, a two-stage foreground-guided litchi counting framework. In the first stage, an enhanced fruit-cluster detector improves the localization of small and ambiguous clusters under complex canopy backgrounds. In the second stage, the detected foreground regions are fed into a density-regression network with hybrid attention, while a consistency-based training strategy is introduced to improve robustness to appearance and illumination variations. To support this study, a large-scale litchi counting dataset was established, consisting of 1,126 whole-tree images collected from five orchards and spanning three ripening stages, with approximately 120,000 fruit-level dot annotations and more than 20,000 cluster-level bounding boxes. FG-LCNet achieved the best overall counting performance, with an MAE of 7.44 and an RMSE of 11.01. It showed clear advantages in high-density fruit-cluster scenarios and cross-orchard validation, while maintaining competitive results across orchard-region and maturity-stage subsets. The framework further retained inference efficiency suitable for practical deployment. These results indicate that FG-LCNet provides an effective solution for robust litchi counting and offers potential for other clustered fruit-counting tasks.
Reproduction assets foundThe paper's implementation code is explicitly stated as publicly available at the authors' GitHub repository (FG-LCNet). The litchi counting dataset (1,126 whole-tree images with ~120,000 dot annotations and 20,000+ bounding boxes) is not yet fully public: a ~100-image annotated subset is promised upon acceptance, and,Code · publicdustry Technology Research System (CARS-32-21), Hainan Modern Agricul-
655
tural Industry Technology System (HNARS-08-G02).
656
Conflicts of Interest
657
The authors declare that there is no conflict of interest regarding the publication of this article.
658
Data Availability
659
The implementation code is publicly available at https://github.com/johnhamtom/FG-LCNet .
660
Upon acceptance, a representative subset of approximately 100 annotated litchi images will be
661
released to support reproducibility and preliminary benchmarking. The full dataset is being further
662
organized for future release. Before full release, the complete dataset can be obtained from the
663
corresponding authorOpen asset ↗johnhamtom/FG-LCNetpdf-raw-page:28 lines:1-81Plant phenotyping relevance matchOpenAlex · arXiv · checked 15 Sept 2026
Accurate 3D plant organ segmentation is fundamental to automated phenotyping. Existing approaches rely on annotated training data or species-specific model configurations. We present an annotation-free pipeline for 3D plant organ segmentation, combining text-prompted SAM3 segmentation with semantic neural radiance fields (NeRFs). Given only multi-view RGB images and a list of class names, our zero-shot pipeline produces semantically labeled 3D point clouds without manual annotation, per-species fine-tuning, or domain-specific preprocessing. Multi-view NeRF fusion acts as effective implicit consensus mechanism that lifts imperfect per-frame masks into accurate 3D labels. On a controlled Begonia maculata testbed the SAM3 pipeline achieves 92.6% mIoU, reaching 95.9% of the oracle upper bound established with perfect ground-truth masks. The pipeline was further evaluated on a new dataset spanning ten diverse plant point clouds reaching an average 0.856 mIoU, with leaf and pot IoU above 0.91 and 0.90 for every species, respectively. These results demonstrate that annotation-free 3D plant organ segmentation is now feasible and approaching the range of supervised methods.
Medicinal plants are critical sources of bioactive materials, yet efficient field-scale monitoring of their growth responses to water and nitrogen management remains limited. The present work evaluated Unmanned Aerial Vehicle (UAV) multispectral sensing for estimating leaf area index (LAI), underground fresh weight (UFW), aboveground fresh weight (AFW), underground dry weight (UDW), and aboveground dry weight (ADW) in Astragalus membranaceus var. mongholicus (Bunge) P.K.Hsiao ( A. membranaceus ) under 15 irrigation-nitrogen treatments spanning irrigation inputs of 15.0-30.0 mm per event and nitrogen application rates of 0–450 kg N ha⁻¹. UAV imagery and ground measurements were acquired from 45 plots at 60, 75, 90, 105, 120, and 135 days after sowing (DAS). Thirty-seven vegetation indices (VIs) were screened, and random forest (RF), support vector machine (SVM), back-propagation neural network (BP), genetic algorithm-optimized BP neural network (GA-BP), and decision-level fusion (DLF) models were evaluated with repeated random partitions. Growth responses were trait-dependent. W1N5 reached the peak final LAI (1.96), whereas W2N2 achieved the maximum UFW, AFW, UDW, and ADW (4363.94, 7406.00, 1354.50, and 2596.51 kg ha⁻¹, respectively). VI-trait associations changed with growth stage. GNDVI correlated most closely with LAI at 75 DAS (r = 0.78), and biomass-model performance was stronger during 90–120 DAS than at 60 DAS. DLF was retained in 24 of the 30 stage-specific models. In whole-season modeling, LAI was best estimated by BNDVI input with BP (R 2 = 0.96; RMSE = 0.01), whereas UDW was best estimated by DVI input with DLF (R 2 = 0.93; RMSE = 0.39). Stage-adaptive feature selection increased R 2 by up to 38.60%. Overall, the results indicate that UAV multispectral sensing, stage-adaptive VI screening, and ensemble learning provide a practical framework for monitoring A. membranaceus growth and supporting precision irrigation-nitrogen management in medicinal plant production.
Abstract Maize plant plays crucial role not only in the field of agriculture but also in global economy, since it is the third most cultivated crop across the globe. However, these plants are usually affected by various types of diseases such as blight, common rust, gray leaf spot, etc., protecting the plants from these disease is very important. This research proposes a new deep learning model for disease detection to perform well than other models. The hybrid model uses MobileNetV3 as the backbone architecture integrated with attention, fusion and head. The framework includes data preprocessing, feature extraction, classification and interpretation. We will compare the performance of this model with other models such as VGG16, ResNet50, DenseNet121, ALEXNET, etc. criteria for the final evaluation includes Accuracy, Precision, Recall, F1score, Specificity, Logloss, AUC-ROC curve. Through this proposed model we have achieved an accuracy of 98% which is high than the other models compared. The lightweight nature of MobileNetV3 enables us to implement the model in the mobile and IoT devices also. The present study contributes to the development of deep learning model in the field of agriculture, offering a efficient solution for early maize leaf disease detection.
Reproduction assets foundThe paper's maize leaf disease detection model (MAFH) was trained and evaluated entirely on a public Kaggle image dataset, which the authors explicitly declare in the Data Availability statement. No author code, trained model checkpoints, or other paper-specific assets are stated as publicly available.Dataset · publicig and real-
time datasets and in all the environmental situations.
Funding: This research received no external funding.
Disclosure statement: The authors declare no conflict of interest.
Data Availability
The datasets generated and/or analyzed during the current study are available in the CORN OR
MAIZE LEAF DATASET repository,
https://www.kaggle.com/datasets/smaranjitghose/corn-or-maize-leaf-disease-datasetOpen asset ↗Kaggle · corn-or-maize-leaf-disease-datasetpdf-raw-page:27 lines:1-34Plant phenotyping relevance matchCrossref · checked 15 Sept 2026
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.
Mangrove ecosystems are frequently exposed to high concentrations of iron (Fe) in sediments, resulting in Fe accumulation in plant tissues. Although Fe is an essential micronutrient involved in several metabolic processes, its excess requires efficient mechanisms of compartmentalization and storage to maintain cellular homeostasis. Histochemical detection using the Perls reaction has usually been applied to identify ferric iron (Fe 3+ ) in biological tissues; however, the combination of this technique with ultrastructural and elemental analyses remains relatively unexplored in plant cells. In this study, we investigated Fe localization in leaf tissues of Rhizophora mangle L. (Rhizophoraceae), a dominant mangrove species, by combining complementary approaches, including Perls cytochemical reaction, transmission electron microscopy (TEM), scanning transmission electron microscopy coupled with high-angle annular dark-field imaging (STEM-HAADF), and energy-dispersive X-ray spectroscopy (EDS). Perls-positive electron-dense deposits were visualized at the ultrastructural level, and their elemental composition was further characterized by EDS analyses. Fe-containing deposits were detected in the epidermis, mesophyll parenchyma, mucilage cells, and vascular tissues, as well as in multiple cellular compartments, including plastids, mitochondria, vacuoles, cell walls, intercellular spaces, and plasmodesmata, whereas sclerenchyma cells showed no detectable Fe-containing deposits. The combination of Perls reaction with TEM, STEM-HAADF, and EDS provides a complementary approach for high-resolution visualization and elemental characterization of Fe-containing deposits at the subcellular level. This integrated methodology may facilitate the investigation of Fe distribution and compartmentalization in plant tissues under contrasting conditions of Fe availability.
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.
This study assessed the integration of Unmanned Aerial Vehicles (UAVs) and satellite images (Sentinel-1 and Sentinel-2) in advanced machine-learning techniques to monitor the ABV of quinoa crops (Jacha Grano variety) across the Bolivian Altiplano. The proposed method follows a two-step procedure. First, UAV RGB images were used in photogrammetric and deep-learning (Convolutional Neural Networks-CNN) models to estimate reference quinoa ABVs at a 10 m spatial resolution from the crop canopy 3D model and classification, respectively. Secondly, several spectral and polarization/texture indices derived from Sentinel-2 and -1 images were integrated into three decision-tree-based machine-learning models (Random Forest-RF, Gradient Boosting-GB, eXtreme Gradient Boosting-XGB), and one CNN-based machine-learning model to estimate ABV. Additionally, a Stacking Model (STM) build on top of the three decision-tree-based models was considered for comparison. Model evaluation was also performed in a two-step approach. First, a 10-fold cross-validation strategy was used to highlight ABV sensitivity to Sentinel-2 and Sentinel-1 alone and in combination. Secondly, a Leave-One-Plot-Out Cross-Validation (LOPOCV) strategy was used to avoid autocorrelation between the training and evaluation dataset and therefore provided more insight into ABV mapping potential. The results showed that the combination of Sentinel-1 and Sentinel-2 features in the CNN model achieved the best predictive performance with R2 and RMSE values of 0.64 and 0.39 m3 ∙ 100 m−2, respectively. These findings highlight the potential of integrating multi-source information in advanced artificial intelligence algorithms for quinoa ABV monitoring, offering new insights toward the identification of sustainable practices across remote regions with complex socio-economic contexts.
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.
Abstract Plant diseases substantially reduce global crop yields, and cotton production is particularly vulnerable to field-acquired variability in symptom appearance, background clutter, and illumination changes that limit the reliability and scalability of expert visual inspection. This study aimed to develop an accurate, computationally efficient, and explainable framework for real-time cotton leaf disease recognition that is suitable for deployment on resource-constrained edge devices. Using the SAR-CLD-2024 dataset (322 RGB images captured under natural agricultural conditions across seven categories, including healthy and diseased leaves), images were preprocessed via resizing and normalization and augmented online in the training set (random rotations, flips, brightness/contrast adjustments, and random cropping). An EfficientNet-B3 backbone initialized with ImageNet-pretrained weights was fine-tuned using categorical cross-entropy loss and Adam optimization, with early stopping, checkpointing, regularization, and a fixed-seed 70/15/15 train–validation–test partition to enhance reproducibility and reduce leakage. Performance was evaluated on an independent test set using accuracy, precision, recall, F1-score, MCC, balanced accuracy, Cohen’s kappa, confusion matrix, multi-class ROC/AUC, and precision–recall analysis, alongside computational benchmarking (parameters, FLOPs, memory, and inference latency) and comparative experiments against contemporary CNN, lightweight, and transformer-based models. The model showed stable convergence over 30 epochs with a small training–validation gap, predominantly correct predictions with limited confusion among visually similar classes, consistently high precision–recall behavior under moderate class imbalance, and stable performance across repeated runs with low variability and a tight confidence interval. Grad-CAM heatmaps localized necrotic lesions, discoloration, and infected tissues while largely ignoring background, and failure cases were associated with early-stage symptoms, occlusion, shadows, and inter-class similarity. Overall, the framework provides a reproducible, interpretable, and efficient solution for cotton leaf disease classification with practical implications for trustworthy, low-latency, on-device decision support in precision agriculture.
Reproduction assets foundThe paper's Data Availability statement explicitly names the SAR-CLD-2024 cotton leaf dataset used for all experiments as publicly available on Kaggle with a direct URL. No author analysis code or trained model checkpoint is deposited.Dataset · publicntribute to the development of fully automated, scalable, and real-time smart agriculture
systems.
Declaration
Funding
Datta Meghe Institute of Higher Education and Research Wardha, Maharashtra, India
Data Availability: The SAR-CLD-2024 cotton leaf dataset used in this study is publicly
available through the Kaggle platform at: https://www.kaggle.com/datasets/pantho12/sar-cld-2024-dataset-for-cotton
This dataset includes annotated images of various cotton leaf diseases collected under diverse
environmental conditions. All data utilized in this work are freely accessible, and the data
processing methodology has been described in detail to facilitate reproducibility.
Conflict of interest The aOpen asset ↗Kaggle · SAR-CLD-2024pdf-raw-page:32 lines:1-38Plant phenotyping relevance matchEurope PMC · Crossref · checked 15 Sept 2026
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.
This work presents a gold nanoparticle (Au NP)-boosted disposable paper-based electrochemiluminescence (ECL) biosensor for minimally invasive on-leaf in situ detection of endogenous H 2 O 2 in tomato leaves. The inherent capillary action of filter paper (FP) was employed to simplify reagent delivery, while gold nanoparticles (Au NPs) catalytically activated H 2 O 2 to generate reactive oxygen species, thereby driving the ECL signal output. This simple, low-cost paper-based platform enabled time-resolved monitoring of H 2 O 2 in stressed plants, providing a reliable in situ strategy for evaluating tomato physiology and early disease/pest warning.
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.
Simonetti V, Bonato B, Guerra S, Avesani S, Semenzato L, Bulgheroni M, Gjinaj G, Ravazzolo L, Dadda M, Castiello U.
Growth chamberMultimodalStereoRootStem / branchTrackingGrowth / development / phenology
Understanding plant behaviour requires the integration of multiple phenotypic and physiological signals measured over time under controlled conditions. However, different plant signals are typically studied using separate experimental setups, limiting temporal alignment and integrative analyses.We present Mind(the)Plant, a modular experimental facility designed for the synchronized, long-term acquisition of multimodal plant data, including three-dimensional shoot kinematics, above- and below-ground volatile organic compounds (VOCs) and root imaging. Its modular architecture is designed to accommodate additional acquisition modules, such as electrophysiological signalling, as future extensions. The platform integrates a controlled growth environment with stereovision imaging, high-resolution time-of-flight mass spectrometry and custom rhizocameras. These components are connected through a unified network infrastructure that ensures synchronized acquisition and centralized data handling.We validate the performance of each acquisition module through multi-week recordings, demonstrating high-temporal stability, reliable stereovision synchronization, effective isolation of VOCs signals and robust operation of below-ground imaging. We further illustrate the analytical potential of the platform using a one-day continuous multimodal acquisition combining shoot kinematics, above-ground VOC emissions, rhizocameras observations and environmental data.Mind(the)Plant provides a novel methodological framework for studying plant behaviour, signalling and phenotypic plasticity in ecological and evolutionary research. By enabling coordinated measurements of multiple plant response modalities, the platform supports investigations of dynamic plant-environment and plant-plant interactions from a behavioural perspective.
Reproduction assets foundThe paper's data availability statement explicitly deposits data, code and processing pipelines (supporting the multimodal plant phenotyping measurements and analysis) in a public Zenodo archive with an authors' URL matching an allowed URL.Code · publicf Interest Statement
The authors have no conflicts of interest to declare.
Peer Review
The peer review history for this article is available at https://www.webofscience.com/api/gateway/wos/peer-review/10.1111/2041-210x.70411 .
Data availability Statement
Data, code and processing pipelines supporting this study are available at https://doi.org/10.5281/zenodo.22095454 ( Simonetti & Castiello, 2026 ).
References
Avesani S, Bonato B, Simonetti V, Guerra S, Ravazzolo L, Gjinaj G, Dadda M, Castiello U. Comparing proton transfer reaction (PTR) and adduct ionization mechanism (AIM) for the study of volatile organic compounds. Molecules. 2026;31(3):402. doi: 10.3390/molecules31030402.
Baluška F, LeOpen asset ↗zenodo · 10.5281/zenodo.22095454lines:482-508Plant phenotyping relevance matchOpenAlex · checked 15 Sept 2026
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.
Standardized extraction of quantitative phenotypes from images is increasingly important across plant biology, from ecological and evolutionary studies to genetics, breeding, and functional genomics. However, as large image datasets are increasingly used for trait analysis, many biologically relevant traits, including size, shape, color, and spatial patterning, are still measured manually or using fragmented semi-automated workflows. These limitations reduce throughput, reproducibility, and accessibility, especially for researchers without computational expertise. Here, we present BioIMA, an open-source desktop tool for rapid and standardized phenotyping from biological images. BioIMA integrates foundation model-based segmentation with automated trait computation, allowing users to extract quantitative measurements from images through an intuitive graphical interface and without model training. To validate its performance, we quantified a set of knot morphological traits in two Populus species, as these measurements are typically time-consuming to perform manually. Automatic measurements showed strong agreement with manual ImageJ-based measurements (R2 > 0.95), while reducing per-image processing time by approximately 75% (from ~15 s to ~4 s). BioIMA was further applied to diverse plant datasets, including Helianthus and Rhododendron images with varying morphologies and background conditions. Although developed for plant phenotyping, BioIMA may also be extended to other biological samples where region-based size, shape, or color traits are of interest. By combining accessibility and standardization in a lightweight local application, BioIMA provides a practical community resource for image-based phenotyping in ecological and evolutionary studies.
Reproduction assets foundThe paper's own phenotyping tool BioIMA (source code, documentation, example datasets, and user manual) is publicly available on the authors' GitHub repository, directly supporting the paper's image-based trait extraction and validation analyses.Code · publicis powered by embedded models
97
currently including SAM (Kirillov et al., 2023) and mobile SAM (Zhang et al., 2023),
98
which are executed locally through ONNX Runtime for efficient inference without
99
internet connectivity. Source code, documentation, example datasets, and a user manual
100
are publicly available on GitHub (https://github.com/jingwanglab/BioIMA).101
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for this
this version posted September 3, 2026.
;
https://doi.org/10.64898/2026.08.30.747465
doi:
bioRxiv preprintOpen asset ↗jingwanglab/BioIMApdf-raw-page:4 lines:1-60Plant phenotyping relevance matchOpenAlex · checked 15 Sept 2026
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.
In situ monitoring of plant responses to stress is one of the most challenging aspects of precision agriculture, and the dynamic control of crop growth according to fluctuating environmental factors. Although fluorescence imaging provides a nondestructive approach for monitoring stress-related biomarkers, its performance is often hindered by the low abundance of endogenous signaling molecules and strong tissue autofluorescence. Here, we report a microneedle-integrated sensing platform that enables sensitive detection of endogenous hydrogen peroxide (H 2 O 2 ) in living plants. The platform incorporates a second near-infrared fluorescent nanoprobe composed of Er 3+ -doped lanthanide nanoparticles emitting at 1550 nm and Mo-doped polymetallic oxomolybdates serving as the H 2 O 2 -responsive unit. Embedding the nanoprobe into custom-fabricated microneedles allows precise positioning on plant midribs for continuous monitoring of H 2 O 2 dynamics. Under stress conditions, the system successfully visualized spatiotemporal fluctuations of H 2 O 2 in living tomato, spinach, and tobacco plants. This work establishes a strategy for early stress diagnosis and developing universal plant health monitoring technologies.
Abstract Maize is the staple crop for millions of people in Sub-Saharan Africa, particularly for Zambia. Unfortunately, maize crops are exposed to several serious threats due to their susceptibility to foliar diseases like Maize Rust, Leaf Blight, Leaf Spot, Maize Streak Virus, and Maize Lethal Necrosis that may lead to great yield losses. Conventional methods of crop disease identification consist of field surveys that are not only subjective but also difficult to conduct for smallholder farmers. This paper presents the design and evaluation of a highly optimized version of the EfficientNet-B0 Convolutional Neural Network for the automatic detection of maize leaf diseases using maize leaf images obtained from real-world scenarios. The proposed model utilized the concept of transfer learning with ImageNet pre-trained weights and was trained on the Mendeley Maize Crop Disease (Leaf) Dataset which consists of 30,120 images in nine maize disease classes. The developed fine-tuned EfficientNet-B0 yielded 97.57% classification accuracy, macro precision of 97.61%, macro recall of 97.64%, and macro F1-score of 97.61%. From these results, it is evident that transfer learning and fine-tuning greatly boost maize disease classification accuracy while ensuring high computational efficiency. This study makes a significant contribution to precision agriculture as it offers an accurate and computationally efficient AI-based maize disease classification model, which could help smallholder farmers in early maize disease classification.
Reproduction assets foundThe paper's sole qualifying asset is the public Mendeley Maize Crop Disease (Leaf) Dataset of maize leaf images used for all phenotyping/classification measurements, explicitly declared publicly available with a URL. No author analysis code, trained model checkpoints, or other paper-specific assets are disclosed.Dataset · publiconflicts of interest to publish the paper.
Consent to Publish
All authors have read and approved the final version of the manuscript and agree to its
submission to Discover Networks.
Consent to Participate
Not applicable
Data Availability
The Mendeley Maize Crop Disease (Leaf) Dataset used in this study is publicly available at
https://data.mendeley.com/datasets/6w6gsvghfw
Clinical Trial Number
Not applicable.
Ethics Declaration:
Not applicable.
Competing interests
All authors declare no competing interests.Open asset ↗Mendeley · 6w6gsvghfwpdf-raw-page:55 lines:1-22Plant phenotyping relevance matchCrossref · checked 15 Sept 2026
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.
Occlusion is a major factor limiting accurate three-dimensional (3D) wheat phenotyping. In natural growth conditions, overlapping spikes, leaves, and stems often make only partial target regions visible in single-view images, hindering complete and reliable 3D reconstruction. To address this problem, this study proposes an amodal completion-assisted, sequential framework for single-view 3D reconstruction of occluded wheat. The framework first uses visible prompts to recover the complete appearance and structural cues of occluded targets, and then feeds the completed images into single-view 3D reconstruction models to generate complete 3D structures. We construct the MMWO (Multi-view Multi-instance Wheat Occlusion) dataset from MMW, which is captured under controlled indoor scenarios, by synthesizing diverse occlusion samples through organ-level cutouts, random geometric transformations, and region-constrained pasting, with annotations including visible masks, occlusion masks, and complete target images. Six representative reconstruction methods, including Direct3D, Real3D, SF3D, Spar3D, TRELLIS.2, and Hunyuan3D, are systematically evaluated. Hunyuan3D achieves the best geometric performance, with the lowest mean CD- \(L_1\) and CD- \(L_2\) values of 0.1286 and 0.0536, and the highest mean F-score of 0.5668. SF3D achieves the best rendering quality in terms of PSNR, SSIM, and LPIPS. In addition, Pix2Gestalt completion reduces the estimation errors of spike length, width, and area from 9.31%, 10.89%, and 32.23% to 4.64%, 9.70%, and 9.45%, respectively. These results demonstrate that amodal completion can effectively alleviate occlusion-induced information loss and provide more complete structural priors for single-view 3D wheat reconstruction. This study offers a feasible solution for 3D wheat phenotyping under occlusion and provides a systematic reference for applying 3D generative models to agricultural phenotyping.
Miho H, Barranco Navero D, Trujillo Navas I, Morello Parra P, Koubouris G, Trapero Ramirez C, López-Bernal Á, Oueslati A, Flores Mesas D, García López MT, Valverde Caballero P, Deger RE, Meca E, Perri E, Santilli E, Moral Moral J, Estudillo Cazorla C, Cabello Pozo D, Yousef Yousef M, Acar S, Gurbuz-Veral M, Priego Capote F, Zahiri A, Oulbi S, Márquez Pérez MI, M. Díez C.
Background Olive ( Olea europaea L.) breeding initiatives rely heavily on the extensive and correct characterisation of genetic resources to successfully achieve their goals, such as addressing climate change and emerging diseases challenges. However, the historical lack of standardised phenotyping protocols across multi-environment trials has severely hindered data interoperability and large-scale comparative analyses. Methods Within the Horizon 2020 GEN4OLIVE project, five international olive germplasm banks established a consensus-based methodological framework to systematically evaluate over 500 olive cultivars. We harmonised 14 evaluation protocols covering six fundamental dimensions: phenological and agronomic traits, abiotic stress resilience, biotic stress resilience, olive oil yield and chemical quality, table olive quality assessment, and morphological characterisation and photography. While most protocols were adapted from previously published literature to ensure ease of implementation across different facilities, novel methodologies for frost tolerance and standardised photography were developed de novo. Results The implementation of these consensus methods across five countries proved highly successful. This methodological framework enabled the generation of the largest harmonised, publicly available dataset on olive genetic resources to date, effectively making possible the correct comparation and ranking of the olive cultivars based on their specific characteristics. Conclusions This compendium of methods provides a robust, highly replicable reference point for the standardisation of olive germplasm characterisation and use of shared benchmark cultivars as an effective way for data normalization and comparation. It facilitates future global pre-breeding efforts, ensures international data interoperability, and supports the discovery of resilient cultivars to secure the future of the olive sector.
Reproduction assets foundThe article declares two paper-specific public assets: the GEN4OLIVE phenotypic dataset from evaluating over 500 olive accessions across five germplasm banks, hosted on the project's Olive Varieties Database, and a Zenodo-deposited methodological handbook (Extended Data) containing the 14 protocols, visual assessment, Dataset · publicData and software availability
The phenotypic dataset generated from the evaluation of over 500 olive varieties across the five Mediterranean
germplasm banks using this compendium of protocols and methodologies, is publicly available via the GEN4OLIVE
project repository.
• Repository: GEN4OLIVE Olive Varieties Database.
• Link: https://www.uco.es/ucolivo/gen4olive/olivevarieties (GEN4OLIVE Database, 2025).
Page 8 of 15
Open Research Europe 2026, 6:322 Last updated: 14 SEP 2026Open asset ↗GEN4OLIVE Olive Varieties Databasepdf-raw-page:8 lines:1-44Plant phenotyping relevance matchOpenAlex · checked 15 Sept 2026
Modern crop breeding demands precise organ-level analysis for trait quantification, making plant point cloud segmentation (PPCS) increasingly important. However, conventional deep learning approaches rely heavily on densely annotated datasets that are labor-intensive to acquire. Unified PPCS adaptation from distribution-shifted examples with minimal additional training remains challenging. To address this, we propose PlantC2USeg, a deep transfer learning framework featuring cross-scale consistency learning to explicitly align features across spatial scales and an information-restricted decoding strategy that prevents reconstruction shortcuts and promotes robust adaptation. The resulting pre-training enables stable few-shot generalization across species and sensing conditions, while unified fine-tuning with inherited thresholds further reduces adaptation overhead. Under full supervision on Soybean3D, PlantC2USeg achieves the highest semantic IoU and instance mWCov among compared methods, at 91.91% and 94.62%. With 20 labeled samples, it leads both metrics at 89.78% and 90.27%; with only 10 samples, it retains the highest mWCov of 83.23% while achieving 83.19% IoU. Across HR3D, 10-shot transfer to tobacco, tomato, and sorghum averages 78.41% IoU and 79.42% mWCov, while 22-shot transfer to SYAU-Maize achieves the highest IoU and mRec at 92.75% and 93.51%. Furthermore, a leading category-averaged mIoU of 85.0% on ShapeNet Part demonstrates the framework's capability to handle diverse shape variations beyond agricultural domains. These results demonstrate that PlantC2USeg reduces overall adaptation effort under distribution shifts, enabling scalable plant phenotyping and transferable 3D representation learning beyond agriculture.
To enhance crop performance, intercropping strategies leverage volatile organic compound (VOC)-driven interactions with companion plants that constitutively emit VOCs. Despite the agricultural importance, the mechanisms and kinetics of VOC-mediated sensory transduction in receiver plants eavesdropping on neighbouring non-kin emitters remain largely unknown due to a lack of appropriate non-destructive analytical tools. In this work, we employ multiplexed salicylic acid (SA) and H 2 O 2 nanosensors in Brassica rapa subsp. Chinensis (pak choy) plants to visualize, in real time, reactive oxygen species (ROS) and SA signal transduction following exposure to constitutively-released VOCs from neighbouring aromatic plants-namely sweet basil and spearmint. Unique emitter-specific temporal signatures of ROS and SA were observed in receiver pak choy: sweet basil VOCs induced concomitant generation of ROS and SA at 30 min, whereas spearmint VOCs triggered SA production at 30 min, followed by ROS accumulation. The temporal data enabled the formulation of a diffusion model that quantifies the VOC perception threshold that triggers the distinct early ROS and SA signalling. Transcriptomics analysis at 2 h revealed that both emitters evoke largely distinct changes in pak choy, likely stemming from variations in speed and sequence of the early signal transduction, leading to different phenotypic outcomes. Intercropping with sweet basil led to enhanced pak choy biomass, stress resilience and secondary metabolite accumulation, whereas spearmint as companions had a limited impact. Our study captures in real time, the VOC-induced rapid signalling in receiver plants and its ensuing effect on growth. These nanosensor-enabled findings represent an important advance in deciphering how emitter-specific volatile cues are integrated into plant responses, guiding rational selection of beneficial companion plants for improved yield and nutritional profiles in sustainable agriculture.
Agriculture is being revolutionized through the integration of cutting-edge technologies that support efficiency, productivity, and sustainability. Recent trends in transforming traditional agriculture into smart agricultural systems through the application of Artificial Intelligence (AI), IoT, and drone technology are gaining popularity worldwide. In this context, the current study developed an integrated artificial intelligence-based farming system incorporating Internet of Things sensors, drones, deep learning, and web/mobile applications for timely monitoring of crop and aquatic health. Data were collected using IoT sensors and drones equipped with high-resolution cameras to monitor soil and water parameters, including temperature, humidity, and pH, under various climatic conditions. A dataset of 70,000 images was used for system training, validation, and testing with an 80:10:10 split, together with three months of IoT sensor data. Models including YOLOv8, CNN, Faster R-CNN, ResNet50, MobileNet, EfficientNet, DenseNet, LSTM, and Random Forest were used for pest and disease detection, fish classification, fish disease detection, shrimp disease detection, and monitoring of climatic factors. Data pre-processing included denoising, normalization, missing-value handling, and feature extraction. The designed model showed reliable performance across the tasks, with 88.4% accuracy for shrimp detection, 92.1% for pest detection using YOLOv8, and 93.4% accuracy for plant disease detection using the CNN model. Overall performance was recorded at 97% accuracy, with high precision, F1-score, and mAP, and an RMSE of 1.2 for sensor-based prediction and validation. The current findings indicate the potential of using AI, IoT, and drone technologies to detect biotic and abiotic stresses during farming and support a sustainable agricultural system.
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.
Laiyi Fu · Hongbo Liu · Yanbo Han · Shunkang Ling · Hongming Zhang · Fei-Yue Wang · Danyang Wu · Hequan Sun
MultimodalSegmentation
Crop phenotyping is crucial for advancing plant breeding, yet remains a significant challenge. Manual approaches are labor-intensive and do not scale to the analysis of large datasets, while computational methods like Vision-Language Models (VLMs) lack the adaptability for fine-scale spatial reasoning and diverse phenotyping scenarios. To bridge the gaps, we present iPheno, a fully open-source, domain-specialized multimodal VLM for fine-scale crop phenotyping. A key innovation of iPheno is its dual-aware architecture. First, a spatial-aware feature extractor samples mask regions into local K-Nearest Neighbors (KNN) graphs and enables fine-scale analysis of arbitrary-shaped regions; second, a task-aware Mixture-of-Experts (MoE) routing mechanism activates specialized modules for each phenotyping task. To train iPheno and achieve rigorous benchmarking, we constructed iPheno-120K, a large-scale high-precision dataset designed for multiple phenotyping tasks. Evaluations on iPheno-120k test set and other publicly available datasets showed that iPheno outperformed all fine-tuned baselines, by improving F1-score by 17.1% (LLaVA-1.6-13B) to 28.9% (MiniCPM-o-9B), while achieving the highest inference speed and memory efficiency. A web server (https://ipheno.ai4bread.com), a mobile application (www.ipheno.cn), and a stand-alone PC client (https://github.com/2997029323/iPheno-PC-Client) are available for iPheno.
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
Accurate and efficient monitoring of tea plant growth parameters via remote sensing is essential for precision plantation management. However, spectral indices relying solely on reflectance often exhibit limited sensitivity in capturing complex tea canopy characteristics. This study developed a data-driven framework integrating spectral reflectance, frequency-domain harmonic components, and spatial texture features to construct tri-feature fusion indices (TFIs) and establish machine learning and deep learning models for tea growth monitoring. Ten-band multispectral imagery was acquired using a UAV alongside synchronous field measurements of leaf and plant biomass and nitrogen accumulation. TFIs were constructed through exhaustive feature combinations and optimized via a data-driven search strategy. Subsequently, random forest (RF), multilayer perceptron (MLP), convolutional neural network (CNN), and transformer models were evaluated using a leave-one-site-out cross-validation (LOSO-CV) strategy. The selected TFIs showed strong associations with tea growth parameters within the investigated dataset, with R2 values up to 0.63 and 0.62 for leaf dry matter and leaf nitrogen accumulation, respectively. Models incorporating selected TFIs achieved cross-validated R2 values of 0.56 for leaf dry matter (MLP), 0.59 for plant dry matter (MLP), 0.73 for leaf nitrogen accumulation (MLP), and 0.68 for plant nitrogen accumulation (CNN). These models exhibited competitive predictive performance comparable to RF, although no statistically significant differences in mean absolute error were observed under site-held-out evaluation. Furthermore, model-derived spatial maps provided insights into fine-scale spatial heterogeneity and potential interannual variations in tea growth parameters across representative plantations from 2024 to 2025. Overall, this study provides a UAV-based framework for tea growth parameter estimation by integrating multi-domain information without requiring additional environmental observations.
Fruit quality is a critical determinant of economic returns in pear production, and maintaining an appropriate fruit load (FL) is essential for achieving high yield and quality. As a direct indicator of canopy photosynthetic capacity and assimilate supply, leaf number constitutes the key biological basis of reasonable FL determination under the leaf-to-fruit ratio concept. However, accurate and efficient estimation of leaf number in mature pear trees remains technically challenging, limiting its practical use in precision FL regulation. Here, we propose a data-driven framework for leaf number and reasonable FL estimation by integrating 3D point cloud-derived canopy structure with machine learning. A pipeline for extracting 3D architectural traits was developed and implemented in the software tool FTPCT, enabling rapid and standardized trait acquisition. Through correlation analysis, multicollinearity diagnosis, and variance inflation factor screening, five key traits strongly associated with leaf number were identified and incorporated into five machine learning models optimized using Bayesian optimization. Among them, the optimized random forest regression model achieved the highest and most stable performance, with R 2 of 0.85, RMSE of 239.74, and MAE of 149.26 for test dataset. SHAP analysis identified tree crown volume as the dominant contributor to leaf number estimation. Field validation demonstrated that FL regulation guided by the proposed framework significantly improved fruit weight and size without reducing yield compared with conventional practices. Notably, the proposed approach avoids explicit leaf-level reconstruction and relies on less canopy-scale traits, substantially reducing data requirements and computational cost, and thereby offering strong potential for rapid, field-deployable FL regulation in large-scale orchards.
Reproduction assets foundThe paper's phenotyping analysis assets are the authors' publicly released LeafNumPred source code and trained models, and the FTPCT software for 3D trait extraction from pear tree point clouds. Phenotype/point-cloud datasets are only available on request.Code · public. Supplementary data
The following is the Supplementary data to this article:
Multimedia component 1
mmc1.docx (1.6MB, docx)
Data availability
Data will be made available on request. Anyone who wants to obtain other public data can contact us at taost@njau.edu.cn. The source codes and models have been made publicly available at https://github.com/Zhang-Fanhang/LeafNumPred, and the FTPCT software has been released at https://github.com/Zhang-Fanhang/FTPCT/tree/Installation-package.
References
1.Tao S., Khanizadeh S., Zhang H., Zhang S. Anatomy, ultrastructure and lignin distribution of stone cells in two Pyrus species. Plant Sci. 2009;176:413–419. [Google Scholar]
2.Zhang F., Wang Q., Yuan K.Open asset ↗Zhang-Fanhang/LeafNumPredhtml-lines:284-315Code · publical variations [34,35]. The method for calculating these traits are shown in the Supplementary information 1.
2.5. Software implementation for 3D trait extraction (FTPCT)
To facilitate efficient and standardized extraction of canopy structural traits from point cloud data, we used a standalone software tool, FTPCT (available at: https://github.com/Zhang-Fanhang/FTPCT/tree/Installation-package), which integrates the trait extraction procedures applied in this study. The software provides a graphical user interface, enabling users to process tree-level point cloud data and extract key 3D structural traits without requiring advanced programming skills.
FTPCT implements a series of predefined proOpen asset ↗Zhang-Fanhang/FTPCThtml-lines:138-149Plant phenotyping relevance 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.
Problem: Agriculture plays a pivotal role in the Indian economy, where crop production quality and quantity directly impact the livelihoods of millions. Climate variability, emerging plant diseases, and improper pesticide application contribute significantly to agricultural losses. Early and accurate disease detection is crucial for mitigating crop damage and ensuring food security. Methodology: This study presents a novel deep learning framework based on EfficientNet architecture, enhanced with Progressive Fine-Tuning Strategy for automated plant disease detection. The proposed methodology was evaluated on two benchmark datasets: the Plant Village dataset comprising 20,639 images of tomato, potato, and bell pepper with 15 disease varieties; and a drone-captured rice plant dataset containing 4432 samples from public repositories. The model’s performance was assessed using multiple metrics, including classification accuracy, precision, recall, and F1-score. To strengthen the validation of high accuracy results, additional statistical analyses like class imbalance ratio, Entropy, Chi-Square test, convergence curve, ANOVA test, Tukey’s post hoc HSD test, confidence interval, Cohen’s Kappa result, fold-wise dispersion analysis, mean, std deviation are included in the manuscript. Result: Experimental results demonstrate that the fine-tuned EfficientNetV2-B1 architecture achieved exceptional performance with 99.7% classification accuracy on the PlantVillage dataset and 99.03% accuracy on the drone-based rice disease dataset, significantly outperforming existing state-of-the-art transfer learning models. Model explainability techniques further validated the reliability and interpretability of the predictions, confirming the model’s focus on disease-relevant features.
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
Published1 Sept 2026Journal of the Royal Society, Interface
LiDAR / point cloudFlowerArchitecture / morphology / geometry
The mechanical dynamics of poricidal stamens play a crucial role in buzz pollination, but they have only been studied in a few species showing comparatively less curvature. This paper presents an integrated experimental-computational approach to characterize the dynamic behaviour of the morphologically more complex stamens of Medinilla magnifica. An experimental set-up employing a custom-built shaker and a laser triangulation sensor is developed to measure the stamen motion with high spatial and temporal resolution. The resulting displacement transmissibility reveals three distinct resonances, with the second one falling within the reported frequency range of pollination buzzes. Finite-element and multi-body models are developed and validated against experimental results. The models further reveal how the stamen morphology gives rise to its complex dynamic behaviour. The developed method enables precise parametric modelling of stamens and provides new insight into the mechanical basis of pollen release, thereby offering a framework for studying buzz pollination dynamics across plant taxa.
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.
The automation and standardization of hyperspectral imaging, particularly at high spatial resolution, are essential for advancing plant phenomics and supporting diverse plant science research. We developed HyperBird, a hyperspectral microscopic imaging robot, to automate the acquisition of up to 351 leaf-disc samples in a single tray within approximately 2.4 h and thereby support scalable plant experiments. System calibration established a spatial resolution of 24.3 μ m (full width at half maximum; FWHM), a spectral resolution of 1.78 nm (FWHM), and a depth of field of 4.53 mm, producing hyperspectral data cubes of 2195 × 2000 × 950 pixels across the 400–1000 nm spectral range. System-level and biological-sample consistency assessments confirmed stable spectral measurements during extended scanning sessions and across independently prepared trays. Thermal evaluation confirmed that the 150 W illumination design introduced minimal sample heating during scanning. HyperBird was first evaluated using grape downy and powdery mildews, where no consistent measurable effect of repeated imaging on pathosystem development was detected under the tested experimental conditions. Subsequently, HyperBird was used to characterize spatiotemporal spectral progression in grapevine leaves inoculated with Plasmopara viticola . An automated regional spectral tracing pipeline was developed to isolate spectra from retrospectively traced regions and compare them with whole-leaf averaged spectra across 0–9 days post-inoculation (DPI). Categorical mixed-effects modeling showed that these spatially resolved regional spectra exhibited significant disease-associated spectral change by 5 DPI, with a sharp transition at 6 DPI, whereas whole-leaf spectra showed delayed and weaker disease-aligned changes beginning at 7 DPI. These results demonstrate that HyperBird enables high-throughput, spatially resolved hyperspectral imaging for quantifying localized plant disease progression and provides a scalable platform for studying spectral biology across plant phenotyping applications.
Reproduction assets foundThe paper's Data Availability statement points to a public GitHub repository containing the authors' code and processed data supporting the phenotyping analyses; raw hyperspectral image data are request-only.Code · publicData Availability
857
The code and processed data supporting the findings of this study are available in the GitHub
858
repository at https://github.com/jy773Cornell/HyperBird-Robot. Raw hyperspectral image
859
data are available from the corresponding author upon reasonable request due to file size and storage
860
constraints.
861
Supplementary Materials
862
Supplementary materials accompany this article as a separate document (supplementary.pdf).
863
Supplementary Figure S1. Representative GSAM-based segOpen asset ↗HyperBird-Robotpdf-raw-page:39 lines:1-75Plant phenotyping relevance matchCrossref · checked 14 Sept 2026
Published1 Sept 2026Computers and Electronics in Agriculture
Aboveground biomass density (AGBD) is a key indicator in agricultural systems, directly reflecting crop carbon storage potential and yield levels. The World Food Studies (WOFOST) model is widely used for crop growth simulation due to its process-based interpretability. However, its application is limited by complex calibration demands and struggles with spatial heterogeneity. To address these limitations, this paper proposes an assimilation system that integrates Synthetic Aperture Radar (SAR) data into a modified WOFOST model, which requires simple calibration. WOFOST is run in potential production mode, which assumes ideal conditions to reduce input data requirements. Before assimilation, phenology-related temperature sums are aligned with phenology and meteorological data to match local growth stages. Two quantities are then estimated and updated in the model through data assimilation by minimizing the difference between SAR-derived and simulated AGBD. These quantities are the proposed yield reduction factor, which represents the proportional decrease in potential CO 2 assimilation under stresses, and the initial total dry weight at sowing. Validation experiments were conducted using multi-year cotton datasets from two farms in Georgia, USA, differing in whether irrigation was applied. Compared to WOFOST simulations and evaluated against in situ AGBD measurements, the assimilation results improve agreement and reduce error (approximately 43% RMSE reduction at the rainfed site and 15% at the irrigated site). It also delivers spatial maps of biomass, together with model-derived yield and harvest-index diagnostics, and remains operational under frequent cloud cover where optical observations are sparse. This SAR-based assimilation strategy reduces calibration demands, providing a novel and practical pathway to extend WOFOST applications to diverse agricultural scenarios.
Wu F, Liu S, Zohner CM, Townsend PA, Crowther TW, Peñuelas J, Yang D, Yang N, Dong T, Xu W, Wang Z, Liu X, Dai G, Dong J, Durán SM, Schneider FD, Zeng Y, Cornelissen JHC, Kattge J, Wu J, Asner GP, Cavender-Bares J, Reich PB, Yan Z.
Global trait axes reveal overarching dimensions of plant functional variation. However, how these dimensions are spatially organized within and across forest types remains unclear. We combined drone-based full-range imaging spectroscopy with crown-level measurements of 16 physiological, morphological and biochemical traits across temperate, subtropical and tropical forests in China to enable spatially-explicit trait mapping. Through site-training scenario, leaf-to-canopy scaling and spectral-domain modelling tests, we find that reliable canopy trait retrieval depends not only on trait and spectral coverage, but also on preserving trait-spectral relationships across sites and scales. Spectral predictions recovered observed multivariate covariation, summarizing crown variation into a leaf-economics dimension and two additional biochemical dimensions related to hydro-thermal regulation and defence/metabolism. Mapping these dimensions revealed distinct community-level trait organization alongside substantial species- and crown-level variation within forests. These findings link remotely sensed trait retrieval to environmental filtering and plant functional differentiation, providing a scalable framework for monitoring forest functional diversity.
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.
Whitefly-transmitted begomoviruses cause tomato leaf curl disease (ToLCD). In India, at least 15 begomoviruses are known to cause ToLCD, posing a major challenge to resistance breeding. Although several Ty resistance loci have been introgressed from wild tomato relatives, variable resistance responses are frequently observed, likely due to mixed infections and the absence of a standardized disease scoring system. Moreover, limited knowledge of the infecting begomoviruses in resistant genotypes has hindered the effective use of donor lines in breeding programs. This study evaluated 17 homozygous Ty-gene donor tomato genotypes under natural epiphytotic conditions and identified the associated begomoviruses. To quantify disease severity, a robust disease scoring system was developed using coefficient of infection (CI) by integrating symptom parameters- leaf curling, leaf smalling, stunting, and fruiting, with population-level disease incidence. Field evaluations for two years revealed that genotypes carrying both Ty-2 and Ty-3 loci showed higher resistance, though variability existed among them. Genotypes with Ty-3 alone or Ty-5 + Ty-6 combinations also displayed substantial tolerance, and five genotypes were identified as highly resistant. Molecular indexing revealed frequent mixed infections and identified multiple begomoviruses, including a newly characterized species, tomato leaf curl Ty Pusa virus, alongside tomato leaf curl New Delhi virus, tomato leaf curl Palampur virus, tomato leaf curl Gujarat virus, and tomato leaf curl Joydebpur virus. These findings highlight a shift in begomovirus predominance and possible recombination-driven emergence of new variants. This study provides an integrated framework for evaluating ToLCD resistance and emphasizes the need for continuous reassessment of resistance sources to ensure durable tomato cultivar development.
Year-to-year climate variability poses a challenge for agriculture by increasing crop yield variability; therefore, there is a need to identify genotypes that can withstand these fluctuations. With the right selection criteria, genotypes with yield stability across variable environmental conditions can be selected. Methods such as Finlay-Wilkinson random regression (FWRR) may allow us to use sparse datasets-common in plant breeding pipelines-and incorporate genomic data to leverage phenotypic information from related genotypes to predict yield stability. Our objective was to examine how the number of environments and the variance among those environments affect stability predictions. We also integrate FWRR as a genomic prediction tool for characterizing yield stability, comparing it to the traditional genomic prediction models as a reference. We used three datasets: one highly unbalanced dataset for oats (Avena sativa L.) and two completely balanced datasets with different numbers of environments for barley (Hordeum vulgare L.) and wheat (Triticum aestivum L.). We fit standard Finlay-Wilkinson (FW) and FWRR models to estimate grain yield and stability under various scenarios. We found that the estimated stability values obtained were similar using balanced datasets for FW or FWRR. FWRR also achieved moderate predictive ability for stability using unbalanced datasets under 10-fold cross-validation (CV1) with new genotypes. In terms of environmental representation, selecting the right set of environments for inclusion in the model was more important than adding more environments. Our results suggest the possibility of using FWRR to select stable genotypes earlier in line development, as well as to design resource-efficient stability-testing schemes.
Salt stress can markedly alter seedling architecture, creating a need for non-destructive three-dimensional (3D) phenotyping methods capable of resolving fine plant structures. However, organ-level segmentation of plant point clouds remains challenging because of leaf overlap, slender stems, ambiguous stem–leaf boundaries, and severe class imbalance. In this study, we developed PTV-SegCo, a task-adapted Point Transformer model for organ segmentation and structural phenotyping of coriander seedlings under salt stress. PTV-SegCo integrates efficient channel attention, gated shallow–deep feature fusion, and a combined cross-entropy–Dice loss to improve representation of fine and minority organ structures. The dataset comprised 60 manually annotated 3D point-cloud samples from 12 cultivation trays repeatedly observed over five acquisition dates under six NaCl concentrations (0, 50, 100, 150, 200, and 250 mmol L −1 ). Because the earliest acquisition represented a particularly challenging developmental stage, these 12 samples were used as a fixed early-stage model-selection set, while samples from the remaining four dates were organized into four date-blocked training–validation configurations. Under this internal model-development protocol, PTV-SegCo achieved mean mAcc and mIoU values of 93.05% and 89.05%, respectively, and showed numerically higher performance than its direct backbone PTV-Seg50. These values should be interpreted as internal comparative results rather than as an unbiased estimate of generalization to unseen cultivation trays, and the present results should not be interpreted as establishing the broad competitiveness of PTV-SegCo against other point-based, convolution-based, graph-based, transformer-based, or plant-specific segmentation architectures. After semantic segmentation, reconstructed scenes were metrically calibrated using the known cultivation-tray dimensions, followed by individual-plant separation and quality control. Four reconstruction-derived structural descriptors—plant height, projected area, voxel occupancy volume, and leaf point ratio—were extracted to characterize temporal structural variation under different NaCl treatments. For treatment-level inference, individual-plant measurements were aggregated within each cultivation tray at each acquisition time, with the cultivation tray treated as the independent experimental unit. Independent manual validation showed close agreement for plant height and projected area, with R 2 values of 0.9969 and 0.986, respectively. Overall, the proposed workflow provides a feasible approach for organ-level segmentation and automated 3D structural analysis of small coriander seedlings under salt stress. The extracted descriptors primarily represent reconstruction-derived spatial characteristics and should not be interpreted as direct indicators of physiological status; voxel occupancy volume and leaf point ratio remain without direct external validation.
Abstract Tree crowns are complex, three‐dimensional structures whose morphology varies among species, individuals and environments. Although light detection and ranging (LiDAR) provides high‐resolution, single‐tree point clouds that advance species discrimination and the assessment of intraspecific variation in situ, crown shape is still commonly reduced to low‐dimensional metrics (e.g. crown diameter or crown base height), losing much of its three‐dimensional geometric complexity. We introduce a fully 3D geometric morphometric framework that captures crown shape directly from LiDAR point clouds at both species and individual levels. Pre‐segmented LiDAR single‐tree point clouds of eight temperate forest species were converted into three‐dimensional shape representations using radial bounding volumes (RBVs), which partitioned each crown into a standardized set of vertical layers and radial sectors. Surface points automatically digitized from each RBV formed geospatially aligned, 3D pseudolandmark configurations representing geometric morphometric crown shapes. These configurations served as the input data for multivariate analyses of crown shape variation within and between species. Twelve structural traits, including crown and stem dimensions, were extracted from the same RBVs and integrated into analyses of trait–shape associations. The morphospace of crown shape was structured along different axes of variation in broadleaf species than in conifers. Within these groups, species pairs—such as Fagus versus Quercus and Picea versus Pinus —exhibited contrasting intraspecific morphological gradients, with different structural traits driving shape variation in each. Crown base height and total crown height emerged as the strongest predictors of crown shape. Differences in crown shape among species were primarily captured by symmetric components, with asymmetry providing a negligible signal. Interspecific differentiation was largely driven by architectural variation rather than pure size differences. Morphological differences derived from pseudolandmarks and convolutional neural network features exhibited stronger correlations in conifers than in broadleaf species. We present a reproducible, LiDAR‐native framework for quantifying and comparing 3D crown morphology within and across species. Using the RBV approach, geospatially aligned pseudolandmarks can be derived from any pre‐segmented, single‐tree LiDAR point cloud, enabling scalable, multi‐regional analyses of intraspecific variability. This framework provides a robust foundation for integrating crown shape into ecological, evolutionary, silvicultural and modelling studies, including assessments of environmental effects and architectural constraints.
Maize ear geometry (length, width, curvature, and volume) is closely tied to yield and grain-filling outcomes, but existing high-throughput phenotyping pipelines remain constrained by the cost, labor, and specialized hardware they require. We developed and validated a low-cost pipeline that reconstructs a watertight 3-D mesh of a maize ear from a single 20-second video captured with a consumer-grade DSLR on a motorized turntable under uniform LED illumination. Camera poses from a multi-seed COLMAP procedure initialize a Neural Radiance Field (NeRF), and a cylindrical holder of known diameter, visible in every frame, provides automatic metric scaling with downstream geometric quality control. Applied to 300 ears spanning a diverse maize inbred panel, 250 (83.3%) passed automated processing and quality control. Skeleton length agreed with manual caliper measurements across all 250 ears (R^2 = 0.964, RMSE = 4.68 mm), and convex-hull volume agreed with water-displacement volume on a 15-ear subset spanning the full size range (R^2 = 0.982, RMSE = 5.26 mL). Residual length error grew with ear curvature, whereas bounding-box height, which records the same straight-line chord as calipers, showed no such trend; the discrepancy therefore originates in the measurement definition, since calipers record the chord while skeleton length traces the geodesic arc. The capture hardware costs approximately 607 USD, and operator involvement fell from roughly five minutes to one minute per ear, with all downstream processing running unattended. The platform provides a foundation for breeding-scale 3-D ear phenotyping.
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.
Surface phenotyping underpins plant science, preclinical animal research and entomology, yet across all three the measurement is almost always a photograph, which records a projection and not the surface itself. Here we present the Gentschinator3000 , an open structured-light platform that brings high-end metric surface measurement within reach of laboratories with no optics expertise, combining documented open hardware, open reconstruction software and analysis workflows for under 4000 Euro in components. It resolves a planar reference to 45 µm local flatness, registers full rotations to a loop closure of 156 µm, and performs stably across acquisition ranges that we define. Applying one workflow to a leaf before and after desiccation, to murine anatomy and to a spread lepidopteran, we find that projection underestimates surface area by 11 to 41 %. That error grows with the condition under study, with the evaluation scale and with the direction of view, so it can confound phenotype comparisons dramatically. In murine limbs a 15-degree change of viewing direction shifts a projected inter-segment angle by up to 23.2 degrees, while the three-dimensional angle does not move. Projection geometry can therefore contribute as much to a measured phenotype as the biology it is meant to quantify.
Reproduction assets foundThe paper explicitly deposits three public Zenodo records: reconstructed 3D surfaces of all specimens (including the leaf and hop cone phenotyping measurements), the authors' analysis notebooks with derived and per-panel source data, and the reconstruction software with build documentation and working examples. All areDataset · publicData availability
1149
The reconstructed surfaces supporting this study are available at Zenodo under
1150
https://doi.org/10.5281/zenodo.22167250.54
1151
Source data for all graph panels are provided with this paper; for panels showing
1152
rendered surfaces, the underlying reconstructions are in the same record.
1153
1154
Code availability
1155
The analysis notebooks, environment specifications, derived data and per-panel
1156
source data are available at Zenodo under
1Open asset ↗Zenodo · 10.5281/zenodo.22167250pdf-raw-page:35 lines:1-52Code · public1151
Source data for all graph panels are provided with this paper; for panels showing
1152
rendered surfaces, the underlying reconstructions are in the same record.
1153
1154
Code availability
1155
The analysis notebooks, environment specifications, derived data and per-panel
1156
source data are available at Zenodo under
1157
https://doi.org/10.5281/zenodo.22167598.55
1158
The reconstruction software, build documentation and minimal working examples
1159
are available at https://doi.org/10.5281/zenodo.22167471.56
1160
The software and analysis notebooks are released under the MIT licence and the
1161
hardware design files under CERN-OHL-P v2. The visible-light platform described
1162
hereOpen asset ↗Zenodo · 10.5281/zenodo.22167598pdf-raw-page:35 lines:1-52Plant phenotyping relevance matchCrossref · checked 14 Sept 2026
Published31 Aug 2026International Journal of Advances in Intelligent InformaticsCited by 0 · OpenAlex ↗
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.
Goal. To substantiate the methodical approaches to analyzing the state of soybean crops based on the results of aerial photography with UAV by comparing manual vectorization, controlled classification according to the algorithm of maximum similarity, and uncontrolled classification of K-Means, as well as to determine the feasibility of their combination with expert visual interpretation to assess the spatial structure of the vegetation cover. Methods. Aerial photography of the test proving ground was performed with the help of the unmanned aerial vehicle DJI Phantom 4 Advanced with the subsequent photogrammetric study of materials and the formation of a highly detailed orthophotoplane. In the QGIS environment, visual decryption and manual vectorization of the main objects of the agrolandscape were carried out with the creation of polygonal layers. For automated mapping, methods of controlled classification according to the algorithm of maximum similarity and uncontrolled classification based on the K-Means algorithm were used. The accuracy of the results was evaluated by comparing the data of automated classifications with the data of manual digitization, which was used as a reference (control) method. On the basis of the results obtained, empirical data were summarized to justify practical recommendations for the application of the studied approaches. Results. The study was conducted on the territory of the research farm of the Separate subdivision of the National University of Life and Environmental Sciences of Ukraine «Berezhany Agrotechnical Institute» (vil. Pavliv, Ternopil district, Ternopil oblast) (49.452057°N; 24,805818°E) in may – august 2025. The obtained cartographic materials made it possible to quantify the areas of the main objects of the agro-landscape and identify problem areas with sparse shoots. Methods of controlled classification showed greater compliance with the digitization data (average deviation — 6.7%) compared to uncontrolled (14.7%), which were effective from the point of view of preliminary assessment of spectrally homogeneous sections, but did not provide an accurate division of shoots by density. Analysis of the spatial structure of coverage made it possible to plan local agrotechnical measures and assess the potential yield. Conclusions. Methodical approaches to analyzing the state of soybean crops based on the results of aerial photography with a UAV equipped with RGB cameras are promising and economically feasible. Automated classification methods are effective for highlighting hard and contrasting objects and small-contoured areas, while a detailed assessment of the structure of the vegetation cover is advisable to carry out using a controlled classification in combination with expert visual interpretation.
All living organisms rely on the movement of ions across cell membranes as the fundamental physical basis of their internal energy and signaling, and plants are no exception. Plants perceive, integrate, and respond to environmental stimuli through electrical signals, classified as action, variation, and system potentials, that are coupled with calcium waves, reactive oxygen species, and hydraulic and hormonal changes to coordinate whole-organism responses despite the absence of a nervous system. Yet most studies characterize these signals using a single feature, such as amplitude or spike duration, in a single tissue, an approach that cannot establish how such signals correspond to the underlying ionic activity, mobility, and structural complexity of the signaling environment, or how this correspondence varies across organs. Here, we correlate plant bioelectrical signals with potential ionic energy flow using a multi-domain framework, combining discrete spike events, continuous waveform properties, spectral composition, and signal complexity applied to leaf, stem, and root recordings from tomato ( Solanum lycopersicum ) exposed to different stimulus. Electrical activity with increased stimulus strength, likely reflecting increased ionic flow, with the root showing the largest response. This suggests plant electrical signaling works as a distributed, ion-based information system, useful for stress monitoring and bio-inspired sensor design.
Plant stress biology has traditionally relied on the analysis of snapshot measurements—including hormone levels, reactive oxygen species, transcripts, and physiological traits—which primarily characterize the current state of the cell, whereas the physical consequences of previous stress exposure remain considerably less accessible to direct measurement. This distinction may be particularly important under natural conditions, where plants experience recurrent, sequential, and combined stresses. This raises a fundamental question: can a cell, after physiological recovery, retain a measurable residual physical state that reflects aspects of its previous stress history and influences its response to subsequent stress? Here, we propose Environmental Phase Imprinting (EPI) as a testable biophysical hypothesis according to which environmental stress may leave a measurable imprint on the physical state of biomolecular condensates that persists after cessation of the initial exposure. EPI is not proposed as a new form of biological information or as an established mechanism of stress memory, but rather as a potential physical substrate, correlate, or consequence of previously described forms of cellular stress memory. To operationalize this hypothesis, we introduce the Plant Condensate Code (PCC), a multidimensional conceptual framework designed to move from a static “snapshot” of cellular state toward a time-resolved physical trajectory. PCC integrates complementary characteristics of condensate populations, including morphology, dynamics, molecular mobility, material state, and molecular composition, across a sequence of states encompassing baseline, stress, adaptation, recovery, and the post-stress state.We propose that the trajectory of condensate states, rather than any single measurement, may contain information about cellular stress history and may help explain differences in responses to recurrent or combined stress. Multimodal approaches, including live-cell imaging, fluorescence recovery after photobleaching (FRAP), molecular mobility analysis, microrheology, Brillouin microscopy, quantitative phase imaging, and molecular profiling, could provide complementary measurements of this physical state. PCC is further positioned within our broader conceptual research program encompassing Cytoplasmic Phase Homeostasis, Cytoplasmic Phase Sensing, and the Plant Threat Matrix (PTM)—a proposed six-state framework for describing plant physiological states under stress. Within this framework, physical measurements of condensate and cytoplasmic states may represent one possible approach for defining and quantitatively characterizing cellular physiological states. Finally, we discuss the potential application of this conceptual framework to stress-resilience phenotyping, evaluation of biostimulants, and selection for stress tolerance, while clearly distinguishing experimentally established phenomena from hypotheses and conceptual proposals. The proposed framework may provide a foundation for the development of a new direction in biophysical phenotyping of plant stress resilience, complementing molecular and physiological approaches to the study of stress memory.
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.
Visible-near infrared (Vis-NIR) spectroscopy provides rapid crop disease assessment; however, poor model generalizability remains a major limitation when models developed for a specific period are applied to batches collected at different times, primarily due to variations in physicochemical properties such as chlorophyll content, moisture level, surface texture, and tissue structure, which induce shifts in spectral distributions across batches. This study investigates deep domain adaptation to enhance cross-batch transferability for sugarcane disease classification. Two batches of healthy and symptomatic leaves were collected at different times using the same spectrometer. A customized one-dimensional convolutional neural network (1D-CNN) was trained on Batch 1 and adapted to Batch 2 using labelled samples through two strategies: retraining only the fully connected layers or fine-tuning all network parameters. Both strategies achieved 94% accuracy, with precision 0.92, sensitivity 0.97 and specificity 0.98, outperforming the non-adapted model and standard-free calibration transfer methods, namely Correlation Alignment and Transfer Component Analysis. These findings demonstrate that deep domain adaptation substantially improves the robustness and transferability of Vis-NIR classification models across heterogeneous sampling batches.
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 confirmedOpenAlex · arXiv · checked 5 Sept 2026
High-throughput phenotyping depends on accurate 3D reconstruction of plants across growth stages, yet the development and evaluation of temporal completion methods are limited by the lack of datasets with complete geometric ground truth. To address this challenge, we introduce SynthCrop4D, a procedurally generated synthetic dataset of temporally evolving plant point clouds that provides controllable noise, occlusion, and complete plant geometry for benchmarking reconstruction methods. Using this dataset, we evaluate a two-stage pipeline that combines spatial denoising and temporal point cloud completion. First, a denoising module removes structural artifacts from raw laser-scanned point clouds. The resulting data are then processed by an Adaptive Temporal PoinTr model that reconstructs the current growth stage (t) using information from the previous stage (t-1), enabling recovery of regions missing due to self-occlusion. We evaluate the proposed framework on both SynthCrop4D and the real-world Pheno4D dataset (tomato and maize) under settings with and without denoising. Results show that denoising substantially improves reconstruction quality, with the best configuration achieving a Chamfer Distance of 0.0061 on SynthCrop4D (Temporal PoinTr + Mamba-DG) and an F-Score of 0.2080 on Pheno4D (Vanilla PoinTr + Mamba-DG). We further demonstrate the use of completed point clouds for phenotypic trait extraction, including plant height, canopy width, and convex hull volume, obtaining hull-volume MAEs of 0.021 on synthetic data and 0.343 on real data. Together, SynthCrop4D and the proposed pipeline provide a benchmark and methodology for temporal plant reconstruction and high-throughput crop phenotyping.
Reproduction assets foundThe paper's authors explicitly state that source code, implementation details, and pre-trained model weights are publicly available on GitHub, and that the paper-specific SynthCrop4D synthetic dataset can be reproduced via scripts in that codebase. Pheno4D is a cited prior public dataset, not a paper-specific asset.Code · publicThe source code, implementation details, and pre-trained model weights for this study are publicly available on GitHub at https://github.com/Mrudul2006/3d_plant-reconstruction .Open asset ↗Mrudul2006/3d_plant-reconstructionlines:1724-1761Plant phenotyping relevance matchEurope PMC · OpenAlex · checked 15 Sept 2026
Crossman SG, Wang M, Nowotarski SH, McClain ML, Sankari S.
Alfalfa / lucerneX-ray / CTRoot2D/3D reconstructionVisualization / data managementGrowth / development / phenology
The symbiotic relationship between the legume Medicago sativa and the soil bacteria Sinorhizobium meliloti results in the formation of nitrogen-fixing root nodules. Traditional destructive methods, including paraffin sectioning, vibratome sectioning, and cryosectioning, have been applied to visualize how bacteria occupy the nodule, making it extremely difficult to obtain reliable three-dimensional information. These approaches are often combined with fluorescent labeling or staining, which can introduce additional stress affecting plant growth and nodule formation. MicroCT has emerged as a relatively quick, easy, and robust tool for plant biology that can non-destructively visualize plant histological features in three dimensions (3D), thereby avoiding destructive artifacts during sample preparation and ensuring high-fidelity 3D reconstruction. While microCT has been applied to legume root nodules, a detailed established protocol that documents the process from plant harvest and sample preparation to scanning and software visualization is lacking. In this study, we show a step-by-step microCT workflow using Medicago sativa as a model. The protocol includes nodule excision from roots, fixation, contrast enhancement, mounting, scanning, and three-dimensional reconstruction. Critical parameters affecting elements such as image quality, tissue preservation, and contrast are highlighted. Using this approach, it is possible to visualize the overall tissue organization, bacteroid-infected cells, and vascular bundles in three dimensions without physically sectioning the nodules. The pipeline described here provides a reproducible method for non-destructive, high-resolution imaging of native root nodules and is likely adaptable to other legume species, offering researchers a practical tool for studying nodule structure and bacterial organization within nodules in 3D.
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.
Yaning Zhao · Guodong Zhang · Jiaying Rong · Mingyue Gao · Peipei Dang · Tianjun Zhang · Hao Suo · Zhijun Wang · Ziyong Cheng · Panlai Li · Jun Lin
Raman / spectroscopyObject detectionPhysiological trait estimationWater status / transpiration
ABSTRACT Real‐time, accurate water monitoring is a crucial technical foundation for industrial, environmental, and biological research. However, traditional detection methods typically include destructive processes and suffer from response delays. Near‐infrared (NIR) luminescent metal halides offer a novel solution to this challenge, but they still face issues such as ultraviolet excitation and low photoelectric conversion efficiency. Herein, a luminescent material was synthesized based on the blue‐light‐excited lead‐free perovskite Cs 2 HfCl 6 :Te 4+ /Mo 4+ , in which energy transfer (ET) from Te 4+ to Mo 4+ enables highly efficient NIR luminescence in the 800–1200 nm wavelength range. Upon encapsulation with a commercial blue light chip, the fabricated NIR light‐emitting diode device achieved a photoelectric conversion efficiency of up to 16.1%. By utilizing the absorption characteristics of water molecules in the NIR spectrum and receiving signals via a sensor, an interactive learning process based on a neural network machine learning algorithm was employed, achieving an estimation accuracy of up to 98.6% for plant water content. This non‐destructive and precise NIR detection module provides a new solution for the real‐time monitoring of crop physiological status and holds broad application prospects in the fields of precision agriculture and plant science.
The increasing world population necessitates new sustainable nutrient sources, making microalgae like Chlorella sorokiniana interesting due to its rich nutrient profile and sustainable cultivation methods. With genetic optimization tools like CRISPR/Cas9, microalgae as a nutrient source can be improved even further. However, degradation of the rigid cell wall of microalgae, and thereby developing protoplasts, is often necessary prior to transformation, but monitoring protoplast development in spherical, single-celled organisms like C. sorokiniana is challenging using bright-field microscopy. Carbotrace 480 and 630 were tested as fluorescent markers of the cell wall of a C. sorokiniana mutant for protoplast detection, and Carbotrace 480 was successfully used to distinguish protoplast from normal cells in a cell suspension. The enzymes Driselase, Glucanex, Snailase, and Saczyme were tested in different combinations to degrade the cell wall of the mutant, with Snailase as the most effective yielding ~60 % protoplasts. This study provides a quick and easy tool for monitoring protoplast development in the microalgae C. sorokiniana, the first step to improve C. sorokiniana as a sustainable nutrient source using genetic optimization tools like CRISPR/Cas9.
Main conclusion We established a protocol for reliable nuclear visualization in Charophyceae, revealed diverse nuclear organization across cell types and species, and identified suitable cells for genome size estimation via flow cytometry. Charophyceae are multicellular green algae closely related to land plants and are established model systems for understanding plant evolution. Yet key cellular parameters like genome size remain poorly characterized. We combined fluorescence microscopy, transmission electron microscopy (TEM), and flow cytometry to characterize nuclear diversity across cell types and species of Characeae and to identify a cell type suitable for genome size estimation. Among three DNA-intercalating fluorochromes, propidium iodide labeled nuclei most reliably. Nuclear morphology varied widely across cell types: mononucleated cells were found in vegetative apical cells, the coronula of oogonia, and spermatogenous filaments of antheridia, whereas multinucleation predominated in other tissues, e.g., cortical cells, spine cells, stipulodes and rhizoids. Nuclei in Chara hispida showed a significant gradient in cross-sectional area along the thallus axis. In the apical internodes, nuclei were larger and more heterogeneous, whereas in the basal internodes they were smaller and more uniform, which is consistent with possible endopolyploidy. TEM confirmed the nuclear identity of crescent-shaped structures. Relatively large nuclei were found in rhizoids and spine cells. Only Sphaerochara canadensis showed an organized nuclei pattern. Whole-thallus preparations did not yield a defined nuclear peak by flow cytometry, but antheridia of Chara tomentosa produced a sharp peak, from which a genome size of 5.32 pg (1C) was estimated. The protocol established here provides a simple, reproducible framework for visualizing nuclei and estimating genome size in Charophyceae, and helps address longstanding questions in this group, such as the mechanisms and functions of multinucleation, the site of meiosis, and genome evolution.
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.
While most phenotyping platforms rely primarily on image-based measurements, advanced plant characterization requires the integration of active physiological sensing modali- ties such as chlorophyll fluorescence. We present an autonomous robotic platform designed to perform targeted fluorescence measurements on plant leaves. The system combines 3D plant reconstruction, geometric analysis, and motion planning to localize suitable measurement points and generate collision-free trajectories for a robotic manipulator. A dense 3D model of the plant is reconstructed from multi-view data and used to extract candidate leaf surfaces based on orientation, accessibility, and sensing constraints. These targets are then integrated into a task-level planning framework that guides the end-effector to precise contact or near-contact configurations required for point-based fluorescence acquisition. The platform enables automated, repeatable, and spatially resolved physiological measurements that go beyond passive imaging. By tightly coupling perception, geometric reasoning, and manipulation, the proposed system provides a robotics-driven approach to high-resolution plant phenotyping and opens new directions for autonomous agricultural inspection and plant-aware manipulation.
Reproduction assets foundThe paper states its code is publicly available in the authors' SonyCSLParis GitHub repository (Plant3DImager), which implements the phenotyping perception and motion-planning pipeline. The exact full URL is split across a line break in the supplied text, so the verifiable allowed URL prefix is used.Code · public2 The code is available at https://github.com/SonyCSLParis/ 3 See for example at https://www.youtube.com/watch?v=Open asset ↗SonyCSLParis/pdf-page:4 lines:1-61Plant phenotyping relevance matchOpenAlex · checked 11 Sept 2026
Accurate quantification of rice aboveground biomass (AGB) is critical for crop monitoring but remains challenging due to the complex nonlinearity arising from the coupling of plant density, spatial structure, and internal dry matter distribution. To address the limitations of single-source remote sensing, this study proposed a Three-Dimensional Dry Matter Distribution Integration (3D-DMI) model, which establishes a physically interpretable framework decomposing AGB into dry matter density ( ρ ), horizontal projection distribution ( S ), and vertical cumulative distribution ( h d ) components. Guided by this framework, a core subset of six features (Red_650, MTCI, G_correlation, R_correlation, LPI, and HPA0_99) was extracted from UAV-based multispectral, RGB, and LiDAR data using a dual-step feature selection approach combining Maximum Information Coefficient (MIC) and Distance Correlation (dCor). A Random Forest (RF) regression model was then developed to estimate AGB across the entire growth season. The results demonstrated that the 3D-DMI model achieved excellent performance with an R 2 of 0.920, an RMSE of 0.184 kg/m², and an RPD of 3.544, significantly outperforming any single-sensor approach. Single-feature analysis revealed that while LiDAR-derived structural features provided the fundamental basis for biomass estimation, they encountered inherent saturation bottlenecks during late growth stages. Feature contribution analysis based on SHAP further quantified that LiDAR-derived features dominated the estimation process (68.5% contribution), providing the volumetric basis, whereas RGB textures (18.3%) and multispectral features (13.3%) provided indispensable supplements. Ultimately, this study established a robust, physically grounded computational paradigm for high-precision UAV-based rice biomass monitoring across the entire growth cycle.
Sebastian Varela · Jeremy Ruhter · Erik Sacks · Xu-ying Zheng · Dylan Allen · Anna Hale · Cory Landry · Xian-yan Kuang · Blake Long · Ernst Cebert · Yan Zhu · Shatabdi D. Proma · Sehijpreet Kaur · Diego Jarquin · Jesse Morrison · Andrew D. B. Leakey
Abstract The integration of digital technologies for high-throughput field phenotyping is critical for accelerating crop improvement in agriculture. However, extracting traits from remote sensing data remains constrained by fragmented workflows, manual intervention, and limited interoperability among existing tools, resulting in delays that hinder timely biological insight and decision-making. To address these challenges, we present PhenoStream (Phenotyping Streaming), a scalable, end-to-end cyberinfrastructure designed to automate the full lifecycle of aerial imagery-based phenotyping, from data acquisition to plot- and genotype-level inference. The framework integrates automated data ingestion from distributed field sites, geospatial processing, and AI-enabled trait extraction within a unified, user-accessible graphical interface. Its modular and extensible architecture supports adaptable trait modeling and seamless integration of new data sources, enabling deployment across diverse crops, environments, and experimental designs. We demonstrate the system across a large multi-location field trial network of bioenergy crops, where it enables high-throughput characterization of spatiotemporal growth dynamics, genotype-by-environment (G×E) interactions, and predictive modeling of key agronomic traits. By significantly reducing processing latency and manual effort, the platform facilitates near-real-time analysis and reproducible workflows. This work establishes a generalizable and scalable pathway for operationalizing very-high-spatial resolution aerial phenotyping in agricultural research. By bridging data acquisition and analytics, the end-to-end cyberinfrastructure provides a foundation for integrating heterogeneous and unstructured data streams-including remote sensing, environmental, and management data - toward data-driven decision making in agriculture.
Chlorophyll fluorescence provides sensitive information on plant photochemical responses, but its measurement requirements can limit high-throughput application. This study investigated whether RGB imagery could approximate chlorophyll-fluorescence-derived photochemical status across garden plant species during progressive soil drying. A Photochemical Status Index (PSI) was constructed by principal component analysis from five highly correlated JIP-test energy-flux variables (RC/CS, ABS/CS, TRo/CS, ET2o/CS, and RE1o/CS). The dataset comprised 50 aggregated species-by-soil-moisture-stage observations representing ten species and five sequential soil-moisture stages. Eleven RGB-derived variables were evaluated, and a partial least-squares regression model was assessed using nested leave-one-species-out validation, with all data-dependent procedures repeated within each outer training fold. PC1 explained 96.5% of the shared variation among the fluorescence-derived fluxes. The predictors g, GLI, ExG, ExGR, and CIVE were retained in all ten outer folds. The final model yielded a pooled out-of-fold R2 of 0.469, an RMSE of 1.585, and an MAE of 1.183. However, species-specific R2 ranged from −0.179 to 0.959, and a calibration slope of 0.509 indicated prediction-range compression. These findings provide proof-of-concept evidence of moderate RGB-based approximation of fluorescence-derived photochemical status, but inconsistent species transferability and the common soil-moisture/time gradient require external validation before practical deployment.
Agricultural operations such as pepper harvesting, fruit counting, and field phenotyping rely on accurate visual recognition and instance segmentation algorithms. However, pepper fruits in complex field environments often exhibit slender and curved shapes, partial occlusion, ambiguous boundaries, and adhesion between adjacent instances. Existing object detection and instance segmentation methods therefore struggle to obtain complete fruit masks, which adversely affects subsequent fruit counting, contour measurement, and picking-point localization. To improve the instance segmentation accuracy of occluded peppers in complex field scenes, this study proposes OccPepSeg-YOLO, an improved model based on YOLO11n-seg. First, a P2FreqFusion module is introduced to fuse shallow, high-resolution detail features with deep semantic features, thereby enhancing the representation of fruit edges and tip regions. Second, an ASC module is designed to model the directional and scale-related morphological characteristics of pepper fruits, while a BoundaryGate module strengthens responses at occlusion interfaces and boundaries between adjacent instances. Finally, an OccPepSegment multi-scale prototype segmentation head is constructed, and a BDoU loss function is introduced to improve the boundary consistency of instance masks. Experiments on a self-constructed field-pepper instance segmentation dataset showed that OccPepSeg-YOLO achieved M-P, M-R, M-mAP50, and M-mAP50–95 values of 93.87%, 92.09%, 97.17%, and 82.31%, respectively, representing improvements of 5.59, 3.18, 3.83, and 9.52 percentage points over YOLO11n-seg. Further comparisons with representative YOLO-based instance segmentation models, including YOLOv8n-seg, YOLOv9c-seg, YOLO12n-seg, and YOLOv26n-seg, demonstrated that OccPepSeg-YOLO achieved the best overall segmentation performance. In particular, its M-mAP50–95 exceeded the best competing result obtained by YOLOv9c-seg by 8.35 percentage points. Under a unified repeated-inference protocol on an RTX 3090 GPU using FP32 precision, a batch size of 1, and 640 × 640 inputs, OccPepSeg-YOLO achieved a mean inference latency of 15.801 ± 1.238 ms, a P95 latency of 17.323 ms, and a throughput of 63.29 FPS. These results demonstrate that the proposed model can produce more complete pepper instance masks under leaf occlusion, fruit overlap, and complex background conditions, providing technical support for field-pepper recognition, fruit counting, and visual perception by agricultural robots.
High-resolution monitoring of forest structure and productivity is essential for effective natural resource management. However, monitoring approaches such as field-based forest inventories or extensive lidar campaigns are costly, time-intensive, and spatially limited. Therefore, inexpensive and accessible methods are needed. SatCHM (Satellite Canopy Height Model) was developed to be an accessible and open-source tool for researchers, allowing for site-specific and temporally flexible predictions of canopy height with limited computational resources. SatCHM requires four inputs: panchromatic satellite imagery, solar and sensor angle metadata of satellite imagery, digital elevation models (DEMs), and lidar-produced CHMs for an area of interest. After SatCHM pre-processes inputs, data is loaded into a collection of convolutional neural networks (CNNs) for image-to-image regression. This ensemble cooperates to yield high-resolution predictions (up to 0.5-meter) of three-dimensional tree structure with discernible tree crowns across a broader defined area of interest. After calculating the mean absolute error for each prediction output, the median of these mean absolute errors was 6.06 meters.
ABSTRACT Leaf shape is a fundamental trait of plant ecological strategies, influencing biotic interactions and ecosystem functioning. However, established quantitative metrics fail to capture subtle variations and irregularities, require user-based reference points or are challenging to compare among taxa with broadly different leaf shapes. In addition, established metrics typically conflate (aggregate) leaf edge complexity and macro-shape complexity, despite their independent functional significance and genetic foundations. Here, we introduce an entropy-based framework to quantify two new complexity metrics: edge complexity and macro-shape complexity. Based on three case studies, we show that these metrics outperform aggregate metrics in predicting Quercus robur chemical traits, provide more intuitive interspecific classifications, and strongly align with human perception. In addition, edge and macro-shape complexity show high complementarity, while aggregate metrics are highly redundant and typically strongly related to leaf area. Emerging as the strongest predictor of leaf chemistry and key visual cue for complexity as perceived by humans, the effects of edge complexity highlight the under-appreciated functional significance of leaf margins. Our framework and the proposed entropy-based complexity metrics thus promise to help unlock the potential of growing digital image archives of leaves, including images from herbaria and fossils, and are technically readily applicable to shapes of algae, bacteria, pollen, and beyond. The accompanying package ShapeComplexity enables the broad application of entropy-based metrics, providing a powerful tool to explore how the shape of organisms and biological structures influences ecological strategies, biotic interactions, and ecosystem functioning while tracking spatial and temporal variation.
Reproduction assets foundThe paper's authors publicly release their ShapeComplexity analysis code (Rust) on GitHub, used to compute the paper's leaf edge- and macro-shape complexity metrics. Supplementary data/analysis code are on Dryad, but that URL is not in the allowed list. RMBG is a generic third-party background-removal model, not a phenCode · publicThe complete, open-source Rust-code (The Rust Team, 2025 ) is publicly available on GitHub ( https://github.com/Thornbach/ShapeComplexity ), ensuring transparency and reproducibilityOpen asset ↗Thornbach/ShapeComplexitylines:86-94Plant phenotyping relevance matchEurope PMC · checked 15 Sept 2026
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.
The intrinsic properties of plants offer numerous opportunities for scientific and technological advancement. Considerable efforts have been directed toward developing plant-on-chip platforms to investigate cellular responses to external stimuli, including chemical, mechanical, and electrical cues. In this study, we present a fluidic platform using polydimethylsiloxane (PDMS) and a printed circuit board (PCB), integrated with electrochemical impedance spectroscopy (EIS) detection. Various experimental conditions were examined, including ionic and pH stimulation, as well as membrane dimensions, with the onion inner membrane treated as a black-box system. The measurement results are presented as Nyquist plots, and a resistance model incorporating multifactorial influences is proposed. Impedance variations in plant cells serve as a basis for electrical modulation. To explore these properties, we converted acoustic signals into electrical inputs and recorded the outputs after being modulated by onion inner epidermal cells. A transfer function analysis was subsequently performed. Our results indicate that the plant cell-on-chip (PCOC) platform holds promise for further investigations into plant cell properties. The impedance results suggest that plant cells can respond to different external stimuli, enabling modulation of the electrical properties. These findings lay the groundwork for future studies on cellular electrical characteristics and the development of preliminary bioelectrical circuits.
Maize diseases affecting leaves, stalks, and ears can substantially reduce yield and quality; therefore, rapid and accurate recognition in complex field environments is important for intelligent agricultural monitoring. To address the large-scale variation, weak fine-grained texture, and strong background interference associated with multi-part maize diseases, this study proposes YOLOv11-MPD (YOLOv11 for Maize Multi-Part Disease Detection), a maize disease detection algorithm based on YOLOv11n. The method jointly improves spatial position awareness, shallow detail preservation, local-context modeling, key semantic-region enhancement, and lightweight detection-head reconstruction. RFCAConv, C3k2_RFCAConv, and Detect_LSDECD are introduced into the baseline network to strengthen directional texture modeling, multi-scale feature aggregation, and detection-head feature representation. FG-RFCAConv, HGD-C3k2, LCA-C3k2, and GRN-BiAttn are further designed for high-frequency differential gated detail compensation, P3 high-resolution detail enhancement, local-context fusion, and global-response-normalized attention regulation, respectively. Experimental results show that YOLOv11-MPD achieves Precision, Recall, mAP50, and mAP50-95 of 72.3%, 72.8%, 79.5%, and 50.2%, improving YOLOv11n by 2.4, 2.5, 2.9, and 2.4 percentage points, respectively, while reducing parameters from 2.6 M to 2.4 M. These results indicate that, within the scope of the dataset used in this study, YOLOv11-MPD improves multi-part maize disease detection under complex field conditions. However, the current conclusions are limited to the constructed dataset, and further validation using larger multi-region, multi-season, and multi-device datasets is required to evaluate its broader generalization ability.
Lodging in sorghum presents a significant challenge for plant breeders due to the trade-off between lodging resistance and grain yield. Manually measuring lodging across thousands of plots is time-consuming, expensive, and error-prone, making selection for lodging resistance challenging in breeding programs. Unmanned aerial vehicle (UAV)-derived metrics provide a potential high-throughput alternative; however, it remains unclear whether photogrammetric heights derived from UAV imagery can estimate plot-level lodging severity in large sorghum breeding trials. This study developed a framework for predicting plot-level lodging from UAV imagery across 2,675 sorghum breeding plots. Multi-temporal canopy height data were collected at two critical time points: maximum crop height and at manual lodging assessment. Height percentiles were extracted from UAV-derived point clouds generated using photogrammetric algorithms. These data were used to develop parametric, non-parametric, and ensemble prediction models, which were evaluated using three statistical metrics. The ensemble model, averaging predictions from all models, achieved the highest accuracy with Pearson correlations of r = 0.80-0.84 and lowest root mean square error (RMSE=16-18%), explaining 64-70% of variation in manual lodging counts. Model diagnostics and iterative refinement, including inspection of UAV imagery and dataset curation, had minimal impact on model performance, demonstrating the robustness of the approach. Model performance was consistent across sites, with minimal effects of stratified sampling on accuracy, confirming the ensemble approach as optimal for plot-level lodging assessment. This study demonstrates that integrated multi-temporal UAV imagery offers a practical alternative to labor-intensive manual evaluation methods by enabling high-throughput lodging assessment suitable for implementation in sorghum breeding programs.
Abstract Deep convolutional networks now classify leaf images on curated benchmarks such as PlantVillage with accuracies close to the measurement ceiling of those datasets, which has shifted the open questions away from raw accuracy toward two under-reported issues: which components of a composite pipeline actually cause the result, and whether the components that matter on laboratory images are the same ones that matter on field photographs. We address both with a classification framework evaluated under a protocol that fixes every development decision before the independent test data are read. A Mask R-CNN stage localizes the dominant leaf, the accepted box is expanded by a validation-selected 5% margin and resized to a shared 224 by 224 input, and the same localized image is passed to ResNet50, InceptionV3, and MobileNetV2 together with an extractor that produces eighteen colour and shape descriptors. The forty-five class probabilities and eighteen descriptors form a sixty-three-dimensional input to a neural meta-classifier developed by three repetitions of stratified five-fold cross-validation. On a locked 3,101-image PlantVillage test partition of fifteen classes the framework reached 99.77% accuracy and 99.75% macro F1 with seven misclassifications, and a one-component-at-a-time ablation confirmed that every stage contributed. The same design was then trained and evaluated entirely within a separate thirteen-class PlantDoc field-image dataset, where it reached 90.03% accuracy and 89.40% macro F1. This is a within-PlantDoc experiment and not a controlled-to-field transfer test, so the figure measures how the pipeline behaves on field imagery rather than how a PlantVillage-trained model survives a domain shift. The central finding comes from running the identical ablation on both datasets: on clean images the ensemble breadth and descriptors provide the incremental gains, but under field conditions the ordering changes, and leaf localization and learned fusion become the decisive components. Removing localization cost 2.83 accuracy points and replacing the neural fusion with soft voting cost a further 2.08 points, the two largest effects on PlantDoc. The contribution is a controlled and transparent account of where each component of a localization-aware plant disease classifier earns its place, and of how that ordering shifts between laboratory and field acquisition.
Reproduction assets found保存済みの本文根拠を更新済みルールで再検証し、公開資産1件を確認しました。Dataset · publicThe classification subset used here (20,638 images spanning fifteen pepper-bell, potato, and tomato classes) was obtained from PlantVillage, which is openly accessible at https://www.kaggle.com/datasets/emmarex/plantdisease.Open asset ↗Kaggle · emmarex/plantdiseaselines:314-336Plant phenotyping relevance matchCrossref · checked 5 Sept 2026
Introduction Accurate estimation of aboveground biomass (AGB) is essential for monitoring pasture productivity and supporting sustainable management of integrated crop–livestock (ICL) systems. We hypothesized that integrating multispectral, thermal, and canopy-structural information derived from unmanned aerial vehicles (UAVs) would improve AGB prediction relative to spectral information alone, and that Generalized Additive Models for Location, Scale and Shape (GAMLSS) would accommodate seasonal heteroscedasticity while maintaining predictive performance comparable to Random Forest (RF) and Support Vector Machine (SVM) models. Methods We collected 280 destructive biomass samples from two ICL paddocks and one continuously grazed pasture in the Brazilian Cerrado between 2022 and 2024. Twenty-four UAV-derived predictors, including spectral bands, vegetation indices, canopy surface temperature, and canopy height, were evaluated using repeated five-fold cross-validation. Model transferability was assessed by withholding one management paddock at a time. Results and discussion Under repeated five-fold cross-validation, GAMLSS achieved the lowest prediction error (R² = 0.69 ± 0.01; RMSE = 2.15 ± 0.04 Mg ha⁻¹), followed closely by SVM (R² = 0.68 ± 0.01; RMSE = 2.19 ± 0.03 Mg ha -1 ); RF showed lower accuracy (R 2 = 0.53 ± 0.01; RMSE = 2.63 ± 0.02 Mg ha -1 ). In the paddock-transferability assessment, GAMLSS also showed the lowest error (R 2 = 0.63 ± 0.04; RMSE = 2.34 ± 0.26 Mg ha -1 ). For GAMLSS, the complete multisensor configuration reduced RMSE by 6.2% compared with the spectral-only configuration. The selected model was used to generate spatially explicit maps of AGB and standing aboveground biomass carbon, estimated from the mean measured carbon concentration of forage biomass. Integrating multispectral, thermal, and structural UAV data with distributional regression improves AGB estimation and enables spatial monitoring of tropical pastures under contrasting management conditions.
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.
Abstract Living tissues contain dynamic biochemical information that is difficult to capture with conventional hyperspectral microscopes because sequential spectral acquisition is poorly matched to in vivo molecular processes that evolve during measurement. Here we introduce a task-specific optical encoding framework for video-rate molecular inference in living plant tissue. The system integrates a passive spectral encoder, implemented here as a low-angle scattering LDPE layer, into a 22-mm miniaturized probe and learns a supervised mapping from ultraviolet-excited autofluorescence measurements to biomolecular abundance maps. Unlike conventional pipelines that first reconstruct hyperspectral datacubes and then perform spectral unmixing, the deployed system directly estimates endogenous molecular contrast associated primarily with lignin and chlorophyll in poplar tissue, with additional suberin-associated contrast evaluated in suberin-rich tissue. This reframing makes the measurement task biomolecular inference rather than spectral reconstruction, enabling biochemical mapping under low-photon autofluorescence conditions while reducing data burden and computational latency. In living poplar stems, the platform captures autofluorescence-derived videos of embolism propagation and wound-induced biochemical remodeling, dynamic processes for which sequential spectral acquisition can introduce temporal mixing because the molecular contrast evolves during the scan itself. The system also resolves genotype-dependent reductions in lignin-associated autofluorescence in engineered poplar lines. Direct molecular inference improves biomolecular estimation relative to a reconstruction-based pipeline, while probabilistic decoding provides uncertainty estimates. These results show that compact passive spectral encoding, when optimized for biological inference rather than datacube recovery, enables deployable, label-free molecular videography of living plant tissue dynamics after task-specific calibration.
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.
Microplastics (MPs) are persistent and ubiquitous contaminants in aquatic ecosystems, yet their interactions with submerged aquatic plants remain poorly understood. While MP-induced phytotoxicity has been extensively investigated in terrestrial plants, quantitative evidence for MP uptake and internal accumulation in submerged species is still limited. In this study, we investigated the phytotoxicity and accumulation patterns of fluorescent microplastics (FMPs) in two submerged aquatic plants, Bacopa lanigera and Rotala indica, using fluorescence spectroscopy. Plants were exposed to FMPs of two particle sizes (50 nm and 1 µm) across three exposure concentrations (0.001%, 0.01%, and 0.05%). Plant growth, chlorophyll content, fluorescence emission, and FMP accumulation were systematically evaluated. Our results demonstrated clear size- and concentration-dependent responses. Smaller particles (50 nm) showed significantly higher uptake and induced stronger phytotoxic effects than 1 µm particles, with pronounced growth inhibition and chlorophyll reduction observed at the highest concentration (0.05%). Fluorescence-based analysis enabled quantitative estimation of both surface-associated and internalized FMPs within plant tissues. Maximum surface accumulation reached 207 ppm, while internal (cross-sectional) accumulation reached up to 75 ppm, regardless of plant species. Under the respective experimental conditions, B. lanigera exhibited higher estimated FMP accumulation, whereas R. indica showed greater growth inhibition. These findings provide quantitative evidence of microplastic uptake and internal accumulation in submerged aquatic plants and highlight particle size as a critical determinant of phytotoxicity. Moreover, this study establishes a fluorescence-based methodological framework for estimating microplastic concentrations in aquatic plant tissues, contributing to improved ecological risk assessment of microplastics in freshwater ecosystems.
Abstract Purpose : The system for diagnosing diseases in Solanaceae crops (SolanAPP), including tomatoes, potatoes, peppers, and eggplants, represents a promising tool for supporting decision-making in agricultural fields using AI. This system reconciles the computational intensity of multitasking models with the infrastructural limitations of rural environments, thereby increasing digital literacy. Its architecture is based on two fundamental pillars: (i) autonomous, offline operation for the detection and classification of diseases in Solanaceae crops; (ii) a georeferenced epidemiological surveillance network with agricultural recommendations for crop monitoring. Methods : The core diagnostic process combines crop-specific semantic segmentation and disease classification models exported to TensorFlow Lite, enabling on-device visual inference and pixel-level severity estimation. An optional online layer integrates Groq’s large language model (LLM)-based reasoning and Firebase services to generate structured agronomic explanations and facilitate the creation of georeferenced community reports when a connection is available. Preprocessing and management of the dataset were performed using the Roboflow platform. The mobile app was developed natively in Kotlin. Model performance was rigorously evaluated using accuracy, recall, F1 score, mean IoU, mPA, inference latency, model size, and decision matrix. The optimal model for the crop was selected using Simple Additive Weighting (SAW). Finally, the overall framework quality and usability were evaluated in Cuba through a user validation survey using a 5-point Likert scale and aligned with the ISO/IEC 25010 software quality model. Results : The model that yielded the best results for most crops was DeepLabV3+ with MobileNetV2, which achieved a classification accuracy of over 97\% while operating with lower inference latency. Beyond individual diagnoses, the system incorporates a collaborative georeferencing feature that allows users to share observations and precise geographic coordinates of detected pathologies to facilitate regional epidemiological monitoring. The user satisfaction survey yielded a satisfaction rating of 4.5/5, with users highlighting the importance of offline diagnosis. Conclusion : Plant disease diagnosis using computer vision can support earlier intervention in resource-constrained agricultural settings, but practical deployment requires models that are accurate, lightweight, interpretable, and usable under limited connectivity. SolanAPP, an offline-first Android framework for detecting foliar pathologies in Solanaceae crops, not only establishes a framework for disease identification in complex natural environments but also provides a theoretical and practical foundation for automated agronomic treatment recommendations and community-based crop surveillance. Impact SolanAPP is a free framework that supports the synergy between multitask deep learning for offline disease diagnosis and LLM-driven reasoning for decision-making in the field. Beyond the quantitative metrics obtained from the selected models, the deployment of SolanAPP in rural contexts serves a fundamental socio-technical purpose: it acts as a catalyst for open access, digital literacy, and agronomic decision-making under unfavorable development conditions. It also represents a strong effort to foster a collaborative epidemiological surveillance network in the agricultural sector. Although it faces challenges, such as the use of field images for model training, this framework marks a promising step in the deployment of edge AI, balancing technical accuracy with practical utility.
Abstract Plant diseases represent major constraints on agricultural productivity, often resulting in significant yield losses. This study presents FitoView, a cloud-based mobile decision support system that integrates deep learning with real-time weather forecasting for sustainable plant disease management. Demonstrated through a case study on Cercospora leaf spot in chili pepper, the system employs custom YOLOv8 models for automated disease detection, classification, and pixel-level severity quantification, combined with meteorological data from the OpenMeteo API. The core innovation lies in an integrated decision matrix that considers three dimensions: AI-assessed disease severity, 48-hour climatic risk forecasts, and optimal spraying conditions, generating four contextualized management scenarios with tailored re-evaluation periods (3-10 days). OpenMeteo API validation across four cities in Sergipe demonstrated very strong correlations, with Pearson coefficients (r) of 0.90-0.97 for temperature, 0.81-0.95 for humidity, and 0.92-0.95 for solar radiation, corresponding to R² values of 0.65-0.94. The YOLOv8 object detection model achieved perfect precision (100%) and macro-averaged recall of 89% across all disease classes, with Cercospora leaf spot detection reaching perfect metrics (100% precision, recall, and F1-score). Field validation in Lagarto, Sergipe, confirmed the system’s practical use: it accurately detected Cercospora leaf spot, estimated severity, and, combined with climatic risk, generated recommendations for alternative treatment and short-term re-evaluation. The Progressive Web Application architecture, deployed on a Cloud Platform, ensures accessibility without installation requirements, while the modular design enables scalability to additional crops and diseases, representing a significant advancement toward democratizing AI-powered precision agriculture tools for smallholder farmers in Brazil.
Maize ( Zea mays L.) is a highly adaptable crop grown worldwide across diverse climates and management practices, with uses across multiple sectors. Consequently, the traits prioritized in maize research and breeding programs vary depending on the specific objectives. Core traits, however, such as flowering time, plant and ear height, and stalk and root lodging, which are important for evaluating and improving the performance and stability of maize genotypes, are routinely evaluated across breeding programs, regardless of their goals. Standardized measurement of these core traits is essential to ensure data reliability and comparability, enabling the integration of phenotypic data across different experiments. Such efforts ultimately support better decision-making and accelerate the development of improved maize genotypes. This is particularly important in public sector programs, where large-scale evaluations, critical for assessing the value of specific genotypes, are often only feasible through collaboration across programs. Here, we provide a protocol for the standardized collection of phenotypic data, specifically focusing on how to measure core traits in maize field trials. These methods promote consistency and accuracy in the evaluation of these traits, and support communication and coordination among groups in the public sector and other research settings. Further, such standardization facilitates the integration and comparison of data across programs, enabling robust longitudinal and multienvironment analyses.
Florian Tanner · Judith Atieno · Sara N. Blake · M. Krysinska-Kaczmarek · Chris Brien · Mohsen Khani · K.R. Clarke · Darren Plett · Jennifer Davidson · Bettina Berger
Abstract Ascochyta blight is a widely occurring chickpea fungal disease that can cause severe yield loss. Breeding for crop resistance benefits from high‐throughput evaluation of plant–pathogen interactions in genotypes which can serve as sources of resistance. Current practice for the evaluation is human visual scoring of disease symptoms, which is limited in throughput and precision. Here, we developed open‐source sensor‐based phenotyping methods using red, green, blue (RGB) and multispectral imaging to measure resistance components and predict disease severity classes in chickpea and wild relatives grown outdoors over three seasons. Pots were imaged at multiple time points with a ground‐based platform, providing 86,792 RGB and 8199 multispectral images. Lesion count was estimated with YOLOv5 (You Only Look Once version 5) object detection (F1 score = 0.27–0.30), fractional green canopy cover was estimated from RGB images, and vegetation indices were extracted from multispectral images. A model trained on growth rates of fractional green canopy cover normalized to control genotypes could predict disease severity classes with an accuracy of 65% –81 % ( 0.43–0.59) on unseen data from three different seasons. The developed methods provide a pathway to predict visual disease severity scores and support the breeding of crops for disease resistance. They may also be used to characterize disease progression, to find underlying resistance mechanisms, and for early disease detection.
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.
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.
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.
Background Global soybean production is constrained by scarce arable land, and standardized evaluation tools remain lacking for natural mixed saline-alkali stress, the predominant abiotic stress under field conditions. Objective This study aimed to establish a comprehensive saline-alkali tolerance evaluation system for soybean germplasms via integrated multivariate statistical methods, and screen core and auxiliary indicators for efficient germplasm identification. Methods Seventy-one soybean germplasms were tested under 90 mmol/L mixed saline-alkali stress (NaCl:Na 2 SO 4 :NaHCO 3 :Na 2 CO 3 = 1:9:9:1, pH 8.2) simulating natural saline-alkali soil. We quantified the saline-alkali tolerance coefficients (SATC) of 13 morphological and physiological indicators, followed by coefficient of variation (CV), principal component analysis (PCA), subordinate function, cluster analysis and regression modeling. Results Significant inter-germplasm variations in saline-alkali tolerance were detected, and indicators with CV > 0.35 ( e.g ., root length (RL), root fresh weight (RFW)) were screened as primary indices. PCA extracted five principal components with 87.18% cumulative variance contribution, and the integrated analytical pipeline categorized germplasms into five tolerance grades: eight highly tolerant, 24 moderately tolerant, 13 generally tolerant, 16 sensitive and 10 highly sensitive accessions. A high-precision prediction model was constructed ( D = 0.290 X 1 - 0.026 X 2 + 0.438 X 3 + 0.402 X 4 + 0.180 X 5 + 0.153 X 6 + 0.813 X 7 - 1.123; R 2 = 0.998, where X 1 - X 7 represent the SATC of germination rate (GR), RL, RFW, total fresh weight (TFW), shoot dry weight (SDW), root dry weight (RDW), and total dry weight (TDW), respectively). A Chi-squared Automatic Interaction Detection (CHAID) decision tree model was further developed and validated using 10-fold cross-validation, yielding a cross-validation risk value of 0.003, which was comparable to the resubstitution risk value (0.002), indicating good generalization ability and low risk of overfitting. A novel five-dimensional overlapping analysis identified RFW as the core evaluation indicator, with RDW, TFW and R/S as key auxiliary indicators. Conclusion This study delivers a standardized, reproducible technical framework for large-scale screening of saline-alkali-tolerant soybean germplasms. It facilitates global saline-alkali land utilization, accelerates worldwide soybean stress-tolerance breeding, and provides a transferable paradigm for stress tolerance evaluation in other major crops.
Abstract Purpose Timely, accurate, and field-scale crop yield mapping is essential for precision crop management, yet most existing studies rely on late- or full-season data, limiting in-season decision-making. This study aims to develop a stage-aware earliest possible corn yield mapping framework that balances early data availability, predictive accuracy, and spatial fidelity by integrating Sentinel-2 imagery, vegetation indices (VIs), and accumulated growing degree days (AGDD). Methods Corn yield data were collected from a commercial farm over three growing seasons (2018–2020). The final modeling dataset included 51,794 Sentinel-2 10 m aggregated yield samples across three seasons: 2018 ( n = 16400), 2019 ( n = 18534), and 2020 ( n = 16860). Sentinel-2 raw spectral bands and derived VIs were organized for V4, V6, R1, R5, and R6 growth stages based on AGDD- and DAP-defined corn growth-stage windows, also confirmed visually on-ground, allowing observations from different planting and harvest dates across the three growing seasons to be aligned by crop developmental stage rather than calendar date. A stage-wise Pearson correlation and frequency-based selection identified informative and non-redundant VI subsets across crop development. Four machine learning models: Random Forest (RF), XGBoost (XGB), k-Nearest Neighbors (kNN), and a Neural Network (NNET; multi-layer perceptron) were trained using 13 input configurations, including individual growth stages and multi-stage combinations capturing phenological progression. Models were tuned via randomized search, trained on 2018–2019 data, and independently validated on the 2020 season. Yield predictions were mapped directly at 10 m Sentinel-2-pixel resolution without spatial interpolation to preserve fine-scale variability. Results Model performance was strongly influenced by phenological stage selection. Among single-stage inputs, R1 was the earliest stage where reliable yield mapping could be availed (RF: R² = 0.56, RMSE = 29.50%). While combining V6 with R1 stage inputs substantially improved predictive performance (RF: R² = 0.70, RMSE = 24.13%) for the yield mapping at the R1 stage, where V6-stage signals provided complementary yield-related information. The full-season combination (V4 + V6 + R1 + R5 + R6) produced the highest accuracy (RF: R² = 0.72, RMSE = 23.28%) but would be less suitable for early in-season decision-making. Early- or late-stage-only inputs (V4, R5, R6) showed weaker and less stable cross-year performance. Among algorithms, RF consistently generalized best to the independent 2020 dataset and is recommended for operational use. XGB showed strong training performance but reduced cross-year stability, kNN yielded moderate accuracy, and NNET achieved accuracy comparable to RF while closely reproducing observed spatial patterns such as center-pivot effects and edge gradients. Direct 10 m mapping preserved yield heterogeneity and avoided smoothing artifacts common in interpolation-based approaches. Conclusion Stage-aware feature selection and phenology-informed input combinations are critical for balancing yield prediction timeliness and accuracy. For operational in-season yield mapping, the RF model using the V6 + R1 stage combination provides a practical and reliable solution, enabling early, accurate, and spatially detailed yield estimates with robust cross-year performance. This study presents a deployable framework for integrating satellite time series and weather data into conventional (non-sequential) machine learning models to support proactive, within-season decision-making in precision agriculture.
Conventional breeding for ideotypes in target environments remains challenging due to genotype-by-environment interactions and the genetic complexity of key agronomic traits. Traditional multi-environment field trials are costly and time-consuming, limiting rapid genetic gain. These challenges highlight the need for digital tools to support rice breeding. However, two major approaches, genomic prediction (GP) and gene-based crop models (GBCMs), have distinct advantages. In this study, a dataset derived from a natural rice population comprising 210 genotypes, genotyped with about 650,000 markers, rice dry matter, and yield across three environments, was used to develop two genomic prediction models, genomic best linear unbiased prediction (GBLUP) and a convolutional neural network (CNN), together with a gene-based crop modeling framework. The effectiveness of these models in predicting rice traits and assisting in breeding selection was subsequently evaluated. Prediction results indicated that biomass and yield could be effectively predicted by all models, with Normalized Root Mean Square Error (NRMSE) values ranging from 10.60% to 18.59% and 9.93% to 18.19%, respectively. In terms of predictive accuracy, parameter-based crop models achieved the highest predictive accuracy, although it was confined to theoretical simulations. This was followed by the GBCM and CNN, whereas the GBLUP exhibited the lowest performance. Furthermore, GGE biplot analysis revealed the predictions of the GBCM aligned more closely with field observations than those of the CNN, emphasizing the potential of GBCM as a practical surrogate for digital breeding. These results provide valuable insights into modeling genotype-by-environment interactions and support the development of data-informed breeding strategies for future rice improvement.
Accurate long-term monitoring of winter wheat phenology is important for crop growth assessment and irrigation management, but 30 m Landsat-based phenology retrieval remains challenging because of sparse observations, cloud contamination and sensor differences. This study developed and evaluated an integrated Landsat-based workflow for monitoring winter wheat phenology in the People’s Victory Canal (PVC) Irrigation Area of northern Henan and the Alar Irrigation Area of southern Xinjiang from 2000 to 2024. Winter wheat areas were mapped using temporally stacked NDVI/EVI features and a CART classifier. Vegetation-index trajectories were reconstructed using locally adjusted cubic-spline capping combined with Savitzky–Golay filtering, and green-up, jointing, heading and maturity were extracted using threshold- and derivative-based detection. The CART-based mapping achieved an overall accuracy of 89.51%, with higher accuracy in Alar (91.45%) than in PVC (84.35%). Compared with S-G-only, Whittaker and TIMESAT-like approaches, LACC + S-G reduced phenological-date errors, especially for green-up and maturity, with RMSE values within 3.1 d against agro-meteorological observations. Phenological stages generally occurred later in Alar than in PVC, and spatial autocorrelation confirmed significant clustering. Agro-meteorological analysis suggested stronger thermal associations in PVC and stronger moisture-related associations in Alar. These results provide practical 30 m phenological information for regional winter wheat monitoring and irrigation scheduling analysis.
Lodging is a major yield-limiting factor in soybean, but efficient large-scale phenotyping and genetic dissection of this complex trait remain challenging for breeding programs. To bridge this gap, this study developed an integrated, breeding-oriented framework that links UAV-based high-throughput phenotyping with candidate gene identification. Field experiments involving 741 diverse soybean genotypes were conducted over two years, with UAV remote sensing performed at key reproductive stages (from R5 to R7). We identified UAV-derived structural (relative plant height), textural (homogeneity, dissimilarity, correlation), and spectral (NDVI, EVI, NDRE) features as the most sensitive indices for retrieving lodging severity. The fusion of these complementary features, coupled with the XGBoost algorithm, achieved high classification accuracy (0.81–0.92) across genotypes, growth stages, and years. This reliable phenotyping pipeline enabled the precise selection of contrasting genotypes (lodging-resistant vs. lodging-prone) for transcriptomic analysis. Transcriptome sequencing revealed 13,447 differentially expressed genes, with significant enrichment in phenylpropanoid and starch–sucrose metabolic pathways. Moreover, the haplotype analysis within a natural population identified superior allelic variants of two candidate genes ( Glyma.19G249100 and Glyma.05G142200 ) significantly associated with soybean lodging resistance. This work can effectively bridge the gap between scalable field phenotyping and the discovery of functionally validated breeding targets, providing an efficient and translational framework to accelerate the development of lodging-resistant soybean varieties.
Canopy 3D architecture is a critical determinant of light interception, photosynthetic efficiency, and final yield in cotton, yet its rapid and accurate characterisation remains challenging in field conditions. To achieve efficient, non-destructive, and quantitative monitoring of the canopy structure of field-grown cotton, this study proposes a 3D structure-based technology stack for high-efficiency, low-cost, and high-precision phenotyping extraction. This stack directly addresses the technical bottlenecks of traditional 3D data acquisition, namely high cost, long processing time, and low operational efficiency, which have hindered large-scale application. We developed a pipeline that integrates a fast reconstruction algorithm with a scale-recovery mechanism using ground control points (GCPs), enabling the generation of true-scale 3D point clouds from UAV aerial images in a cost- and time-effective manner. Using only 141 UAV images and with a reconstruction time of approximately 20 min, we efficiently reconstructed high-quality, scale-accurate point clouds of two 5.5 m × 5.5 m cotton plots, significantly outperforming SfM-MVS and Instant-NGP in terms of both reconstruction efficiency and point cloud completeness. This method, whose current validation is confined to a single season, one growth stage, and two experimental plots, not only achieves a breakthrough by using fewer input images with high efficiency, but also ensures point cloud accuracy and completeness, showing strong potential for rapid field monitoring and real-time management. Based on the high-quality reconstructed point clouds, we further quantitatively evaluated canopy characteristics at harvest, analyzing the coefficient of variation of canopy height, porosity distribution, and canopy volume fraction. The core shortcomings and optimization strategies for the existing canopy structure were identified, providing scientific data support and practical technical references for precision cultivation management and mechanization-compatible planting in cotton.
Accurate estimation of the cotton seedling Leaf Area Index (LAI) is essential for yield prediction and precision crop management. Traditional manual methods are limited by low throughput, while two-dimensional remote sensing techniques often struggle with sparse canopy cover and soil background interference. Although three-dimensional LiDAR presents a promising alternative, existing deep learning approaches typically depend on costly fully supervised point-wise annotations. To address this challenge, this study proposes an end-to-end framework that integrates UAV-based LiDAR, weakly supervised segmentation, and physical parameter inversion. A weakly supervised network, termed CogNet, was developed—incorporating self-distillation and structure-aware label propagation—to achieve precise segmentation of cotton plants using only 10% sparse annotations. Following instance segmentation via Density-Based Spatial Clustering of Applications with Noise (DBSCAN), individual plant phenotypic traits were extracted. A nonlinear Extreme Gradient Boosting (XGBoost) model was then constructed to invert LAI by leveraging allometric relationships between 3D structural parameters and leaf area. Experimental results showed that CogNet achieved an Intersection over Union (IoU) of 85.34%, effectively mitigating overfitting to label noise and achieving performance competitive with the fully supervised RandLA-Net (82.13%). Notably, under the specific conditions of this cotton seedling dataset characterized by strong geometric priors, the weakly supervised model demonstrated enhanced robustness against annotation inconsistencies. The framework attained a plant detection rate of 96.2%, and the XGBoost model delivered high estimation accuracy (R² = 0.879, RMSE = 0.138). This study demonstrates that weakly supervised learning can substantially reduce annotation costs while maintaining model performance, providing an efficient and cost-effective solution for field-scale crop phenotyping.
Early detection of latent fungal decay caused by Penicillium italicum(P. italicum) and Penicillium digitatum(P. digitatum) remains challenging due to the absence of visible symptoms. In this study, a Vis-NIR hyperspectral imaging framework was developed to characterize early biochemical alterations in navel oranges. To address sample scarcity, a generative modeling approach (WGAN-GP) was employed to capture the intrinsic physiological variability of infected tissues. The successive projections algorithm (SPA) identified 20 key wavelengths associated with water redistribution (OH), carbohydrate depletion (CH), and chlorophyll degradation. These wavelengths were expanded into continuous ROI windows (W = 17), enabling integration of narrow-band pigment signals and broad-band absorptions related to water and carbohydrates via a multi-scale mixture-of-experts (MS-MoE) network. The framework achieved a classification accuracy of 97.10% and an F1-score of 0.9666. These results demonstrate that specific spectral absorption windows can serve as reliable, chemically interpretable spectral biomarkers for detecting early pathological changes in citrus fruit.
Sun-induced chlorophyll fluorescence (SIF) is an effective proxy for vegetation photosynthesis, but tower-based retrieval suffers from atmospheric path interference under humid and variable conditions. We present a DOAS-based SIF retrieval algorithm that operates in Fraunhofer lines (680–686 nm, 745–758 nm) and water vapour-sensitive bands (717–727 nm). It constructs an adaptive reference spectrum from SCOPE simulations and PCA and incorporates H2O absorption cross-sections into the fitting process for active atmospheric correction. The algorithm is implemented in a dedicated tower-based system integrating a 1° scanning gimbal with a high-resolution spectrometer. Validation with simulated and field data demonstrates the following: (1) the algorithm retrieves SIF with high fidelity (correlation coefficients >0.9 across all windows); (2) it exhibits lower water-vapour sensitivity and greater cloudy-sky stability than FLD, 3FLD, and SFM, achieving the lowest coefficient of variation (CV = 0.356); (3) over a complete wheat–rice rotation, the retrieved SIF tracks crop growth and phenological stages. This work provides a reliable solution for automated, high-precision tower-based SIF observation under complex atmospheric conditions.
Live imaging of plant subcellular structures is key to deciphering the spatiotemporal bases of cellular processes, and their functional impact on growth and morphogenesis at various biological scales. Live imaging of plant cells essentially relies on expression of fluorescent markers labeling cells or subcellular structures of interest. Simultaneous multi-channel imaging of several markers is still not routine practice in plant cell biology, owing to issues linked to genetic or spectral compatibility of markers, differences in expression levels, silencing, toxicity, etc. Here we designed a three-color marker in Arabidopsis thaliana and Capsella rubella , enabling high-resolution live imaging of plant morphogenesis, including labeling of the cell membrane, the nucleus and the microtubule cytoskeleton. Detection of MT arrays involved the development of a MAP4-MBD-based microtubule marker optimized for plant cells. The three- color marker allows visualization of the three-dimensional organization and dynamics of plant microtubules within the intracellular space with unprecedented precision, in various organs including the root and shoot meristems, the leaf, anther, and gynoecium. Our results demonstrate the potential of such single-construct strategy for cell biology studies in plants.
Accurately quantifying the areas of necrotic lesions on plant leaves is essential for evaluating plant-pathogen interactions and disease resistance. Although digital image analysis methods using ImageJ are widely employed, they often require case-specific optimization and may not be readily applicable across different experimental conditions. Furthermore, many studies have used ImageJ for lesion measurement without providing methodological details, which limits reproducibility. Here, we present a simple, step-by-step ImageJ workflow for measuring irregular necrotic lesions using a standard personal computer and mouse. The procedure relies on manual lesion selection using the freehand selection tool, followed by Gaussian smoothing, binarization, and automated particle analysis to extract lesion area measurements. By balancing manual isolation with computational thresholding, this protocol eliminates the need for extensive parameter tuning. This approach provides an accessible, reliable alternative to time-consuming color thresholding methods, thereby improving transparency and reproducibility in lesion quantification. The workflow's reproducibility has been confirmed through both intra-user and inter-user analyses. Key features • Relies on simple manual freehand selection combined with minimal image processing, requiring only a standard computer and mouse without specialized software or advanced training. • Enables accurate quantification of irregular necrotic lesions in conditions where automated thresholding methods require time-consuming optimization. • Provides a fully detailed, reproducible ImageJ workflow addressing common gaps in published methods, facilitating direct implementation. • Demonstrates high reproducibility, validated by intra-user and inter-user statistical analyses, ensuring reliable lesion quantification regardless of the operator.
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 matchEurope PMC · checked 15 Sept 2026
Fluorescence in situ hybridisation (FISH) is a valuable technique for visualising RNA molecules in their native cellular context. Still, its application in plant tissues is often limited by tissue autofluorescence and the lack of optimised protocols. Here, we developed and validated a simple, reproducible FISH workflow to detect Apple scar skin viroid (ASSVd) in cucumber. Systematic optimisation of probe chemistry, tissue selection, and sampling stage significantly improved assay sensitivity and reproducibility. The AZDye594-labelled antisense riboprobe produced higher signal-to-background ratios and lower background fluorescence than fluorescein-labelled probes, enabling reliable detection of ASSVd in vascular-associated tissues. The optimised workflow consistently detected ASSVd in both leaves and stems. High-resolution confocal imaging further revealed predominant nuclear accumulation of ASSVd RNA in infected cells. Together, this study establishes a sensitive and accessible FISH workflow for localisation of ASSVd in cucumber and provides a practical platform for investigating the spatial distribution of viroid and other plant RNA pathogens.
Andrew Trlica · Rachel L. Cook · Matthew J. Sumnall
Field / plotLiDAR / point cloudMultispectral / hyperspectralLeafMorphology / geometry measurementLeaf traits
Canopy Leaf Area Index (CLAI) is a stand attribute containing information on the real-time health and growth potential of managed pine plantations. Current remote sensing techniques for quantifying CLAI rely on simple linear models applied to satellite multispectral imagery, or on techniques based on light detection and ranging (LiDAR) data that are costly and less frequently collected. This study demonstrates a convolutional neural network (CNN) approach to retrieving CLAI from 10 m Sentinel-2 multispectral imagery with a model trained on gridded LiDAR-based CLAI estimates. We demonstrate large gains in accuracy with the CNN compared to traditional linear models based on vegetation indices (e.g., Simple Ratio), but also clear shortfalls in model skill when predicting “blind” in some spatial domains that were completely excluded during model training. Pixel-scale root mean squared error ranged from 0.34 to 0.64 by domain when exposed to CLAI training data from all available spatial domains, but rose to 0.58–1.74 when predicting without prior domain-specific training. Prediction accuracy was consistently lower when applied to completely unobserved USGS LiDAR-based CLAI estimates. Traditional linear models, in contrast, had the advantage of usually lower prediction error across unobserved spatial domains (0.43–1.98), but with lower maximum accuracy. These results demonstrate a potential route for deploying more complex models for LiDAR “mimicry”, e.g., between data acquisitions widely separated in time, but advocate for the development and use of more stable generalized approaches for use in unobserved managed pine stands.
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.
Wheat grain number integrates fate of individual florets, strongly affected by environmental stress during sensitive stages like meiosis. Because development is asynchronous across tillers, spikelets, and florets, it is hard to distinguish stress tolerance from stress escape. Based on 2400 destructive measurements of spike and anther length together with non-destructive morphological measurements from 158 plants grown in four experiments in controlled-environment, we developed a framework to track individual floret developmental stages at plant level. We applied it in two case studies for connecting within-plant developmental asynchrony to reproductive success under favorable or heat conditions. All florets showed a common relative growth rate, producing additive delays across tillers (1-7 d), spikelets (1-5 d), and floret positions (1-6 d). This generated a developmental map for every floret based on external traits. Under control conditions, grain set probability at floret level combined both positional and developmental effects within a spike. Under heat stress, grain loss occurred only in florets at meiosis during the stress, allowing to quantify a true "stress response", while later florets escaped damage. This framework allows understanding and predicting floret development and linking it to grain set, clearly distinguishing timing effects from positional influences and separating tolerance from stress escape.
Manisha Das Chaity · Ramesh Bhatta · Byron Eng · Jan van Aardt
Field / plotPhotogrammetry / SfM / MVSLiDAR / point cloudMultispectral / hyperspectralWhole plant / canopy / plot / field2D/3D reconstruction
The Greater Cape Floristic Region (GCFR) in South Africa is a fire-prone biodiversity hotspot where high species richness, structural complexity, and small plant sizes (0.0001–4 m2) pose substantial challenges for remote sensing-based biodiversity assessment. Spectral similarity among species and the mismatch between plant size and sensor pixel dimensions limit the capacity of current and forthcoming spaceborne systems to resolve individual species and accurately detect plot-level diversity changes. We therefore developed a physics-based simulation framework that couples fynbos trait measurements with radiative transfer modeling in the DIRSIG (Digital Imaging and Remote Sensing Image Generation) environment towards quantifying information loss across spectral and spatial scales and to define theoretical limits for biodiversity monitoring. We constructed a three-dimensional virtual scene of post-fire fynbos communities in Grootbos Private Nature Reserve, integrating high-resolution imagery, terrestrial laser scanning (TLS), and structure-from-motion (SfM)-derived point clouds. Field measurements of mean diameter and percent cover were used to scale vegetation models and constrain species abundance. We distributed plant instances using a blue noise sampling algorithm, guided by density maps derived from unmanned aerial system (UAS) imagery. Species-specific optical properties were parameterized using field-measured reflectance data and the PROSPECT radiative transfer model, while terrain structure was derived from SfM-based digital terrain models. The integrated scene was used to simulate multispectral (DJI Mavic 3 MSI), hyperspectral (AVIRIS-NG), and light detection and ranging (LiDAR) observations. Agreement between simulated outputs were evaluated against corresponding field-acquired datasets using spectral signatures and vegetation indices. This framework enables systematic assessment of sensor specification effects on spectral biodiversity metrics and provides a pathway for evaluating theoretical limits of species discrimination across airborne and satellite platforms.
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.
Hazra M, Crowther A, McInerney F, Strömberg CAE, Fabillo M.
LeafSeed / grainClassification
Background and aims Phytolith analysis is widely applied in palaeoecological and archaeological research, but its interpretive strength depends on the availability of robust modern reference collections. This study expands the modern Australian phytolith reference collection by analysing 42 native plant species representing 24 families and 37 genera with emphasis on silicification patterns across major growth forms, including forbs, shrubs, trees, and C3 grasses. Methods Phytoliths were extracted from available plant parts, including leaves, stems, flowers, seeds, seed pods, cones, and roots, depending on sample availability. Morphotypes were identified following ICPN 2.0, with grass silica short cell phytoliths (GSSCPs) further classified by shape and size to examine subfamily-level patterns. Phytolith morphotype percentage data were analysed using Hellinger transformation, PerMANOVA, PCA, LDA, and hierarchical clustering to assess compositional differences among plant growth forms and grass subfamilies. Key results Phytolith production varied strongly among growth forms and plant parts. Grasses were abundant producers, whereas most forbs, shrubs, and trees were trace producers or non-producers. Leaves were the most consistent source of phytoliths, while seeds and seed pods were predominantly non-producers. Grass silica short-cell phytolith (GSSCP) morphotypes showed clear subfamily-level differentiation. Pooideae produced Rondel morphotypes. Danthonoideae produced Rondel as well as wide Bilobate types. Panicoideae and Oryzoideae exhibited a pronounced Bilobate signature, commonly associated with Polylobate and Cross forms. Non-grass taxa (woody, shrubs, and forbs) were dominated by Spheroids, Tracheary elements, Epidermal, and Polygonal sheets and other non-diagnostic forms. Phytolith assemblages differed significantly among plant families, with Poaceae uniquely producing GSSCPs, while non-grass families showed greater overlap in assemblage composition. Conclusion By expanding taxonomic and anatomical coverage, this study strengthens the capabilities of phytoliths in the reconstruction of grasslands and in general paleo vegetation in Australia, especially where other proxies such as pollen are limited.
Reproduction assets foundThe authors explicitly state that the R scripts used for data analysis and figure generation are publicly available on their GitHub repository, which directly reproduces this paper's phytolith statistical analyses (PCA, LDA, PerMANOVA, clustering, plots). Supplementary data files contain the paper's measurements but noCode · publicntification of all plant specimens collected for this
study.
A
14 FUNDING
M
15 Funding for this study was provided by ARC Discovery grant DP210100508 and a Ph.D.
D
16 fellowship (UQGSS) to MH.
TE
17 DATA AVAILABILITY
18 The R scripts used for data analysis and figure generation are publicly available on GitHub
EP
19 repository: https://github.com/Manoshi-sporo/Australian-Phytolith-Reference-Collection.
20 CONFLICTS OF INTEREST
CC
21 The authors declare no competing financial or commercial interests.
A
22 AUTHOR CONTRIBUTIONS
23 MH: writing original draft, conceptualization, software, investigation. AC: Supervision, Writing -
24 Review and editing, FM: Supervision, Writing-Review and editing.Open asset ↗Manoshi-sporo/Australian-Phytolith-Reference-Collectionpdf-layout-page:34 lines:1-87Plant phenotyping relevance matchEurope PMC · Crossref · checked 5 Sept 2026
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.
To address the challenges posed by large lesion-scale variations, weak boundary cues, and high inter-class similarity in rice leaf disease detection, this study proposes an improved RT-DETRv2-R50-based rice disease detection model, termed CED-RTDETR. First, a Lightweight Contour-Guided Aggregation Backbone (LCGA-Backbone) is constructed. In the PResNet residual blocks, an InceptionDWConv2d-based direction-aware depthwise separable spatial mixing strategy is introduced to capture local, horizontal, and vertical lesion texture patterns with low computational overhead. Meanwhile, a Contour-guided Efficient Global Aggregation Block (CEGA Block) is embedded after the outputs of the C3, C4, and C5 stages. Through contour-difference enhancement, channel shuffle, group-wise efficient global aggregation, and bottleneck channel mixing, the proposed block strengthens the representation of the boundaries of small lesions, weak textures, and contextual semantics. Second, a Multi-scale Evidence Interaction Fusion Neck (MEIF-Neck) is used to perform multi-scale evidence interaction and salient-region competition after cross-scale feature concatenation. A lightweight feature reconstruction process is further implemented using Spatial-Evidence RepNCSPELAN (SE-RepNCSPELAN), which is developed from a YOLO-style feature fusion structure. Finally, a Direction-Amplitude Decoupled Deformable Attention (DAD-DA) mechanism is introduced to decompose sampling offsets into direction rotation residuals and radius gains while incorporating an aspect-ratio-aware geometric compression-restoration strategy, thereby improving the geometric stability of sampling locations during the decoding stage. Experimental results on the constructed rice leaf disease dataset show that CED-RTDETR achieves AP, AP50, and AP75 values of 29.5%, 75.2%, and 17.6%, respectively, outperforming RT-DETRv2-R50 by 4.1, 4.9, and 3.8 percentage points. To further evaluate the model on an additional public benchmark dataset, experiments were also conducted on the public Rice Disease Dataset. On this dataset, CED-RTDETR achieves AP, AP50, and AP75 values of 34.1%, 77.1%, and 23.2%, respectively, improving upon RT-DETRv2-R50 by 4.3, 5.6, and 3.4 percentage points. These results indicate that the proposed method achieves consistent overall performance improvements on both the constructed dataset and the public benchmark dataset.
Amanda Ewen · Rodrigo Godoy Mendez · Karar Al-Shanoon · Dawn Omoluabi · Anjana Samarasinghe · M. Alejandra Oviedo-Ludena · Karina Chimbo Huatatoca · Kara Glor · Keiko Nabetani · H. R. Kutcher · Lipu Wang · Ian Stavness · Lingling Jin
Reliable and objective phenotyping is essential for plant breeding programs to characterize genetic variation and accelerate crop improvement. Conventional disease assessment relies on expert visual scoring, which is labor-intensive, subjective, and prone to inter- and intra-rater variability. Although image-based phenotyping methods have been proposed, many require manual intervention, specialized imaging setups, or single time-point measurements, limiting their ability to capture disease progression over time. Here, we present a pipeline for longitudinal plant disease phenotyping that quantifies wheat stripe rust and leaf rust progression from time-series images. The pipeline performs semi-automated leaf and automated pustule segmentation from images acquired in situ , enabling objective disease severity estimation with minimal user intervention and without requiring solid backgrounds or manual leaf manipulation or detachment. By extracting temporal traits, including disease severity trajectories and standardized area under the disease progress curve, the method provides a comprehensive characterization of disease development throughout infection. Association between automated and expert assessments was moderate for stripe rust ( R 2 = 0.58) and strong for leaf rust ( R 2 = 0.85), while expert inter-rater reliability was moderate for both diseases (ICC = 0.675 and 0.800, respectively). The proposed approach establishes a scalable and reproducible framework for longitudinal disease phenotyping in controlled environments, with broad applications in disease resistance screening and crop breeding.
Reproduction assets foundThe paper's Code and Data Availability section explicitly states that software and datasets (the phenotyping pipeline and imaging datasets) are publicly available at the authors' GitHub repository and project website, both of which are in the allowed URL list.Code · publicSoftware and datasets are available at:
https://github.com/USask-BINFO/greenskeye_analysis and https://greenskeye.usask.ca/speedbreeding/ .Open asset ↗USask-BINFO/greenskeye_analysislines:195-225Dataset · publicSoftware and datasets are available at:
https://github.com/USask-BINFO/greenskeye_analysis and https://greenskeye.usask.ca/speedbreeding/ .Open asset ↗lines:195-225Plant phenotyping relevance matchCrossref · checked 15 Sept 2026
Localized soil-moisture deficits, that is, irregular sub-field patches where crops experience water stress well before visible wilting, are a leading cause of yield variability in row-crop agriculture. These zones are difficult to detect at the spatial resolution and revisit frequency required for timely irrigation response. This paper presents a reinforcement-learningguided autonomous quadrotor unmanned aerial vehicle (UAV) platform that fuses onboard Visual Simultaneous Localization and Mapping (Visual SLAM) with a pushbroom hyperspectral imaging payload to construct georeferenced, canopy-registered maps of a Crop Water-Stress Index (CWSI) in near real time. Rather than flying a fixed lawnmower survey, the platform is guided by an adaptive-sampling policy trained with Proximal Policy Optimization (PPO) that reallocates flight time and sensor dwell toward regions of emerging water stress as evidence accumulates mid-flight. We present the complete engineering pipeline: airframe and sensor design, a keyframe-based Visual SLAM front and back end that provides centimeter-scale geolocation without continuous reliance on Real-Time Kinematic (RTK) GNSS lock, a hyperspectral preprocessing and spectralindex chain (NDVI, NDRE, NDWI/NDMI) used to derive CWSI through a learned regression, the partially observable Markov Decision Process (POMDP) formulation and reward shaping used to train the sampling policy, and the fused system architecture tying these subsystems together. In simulated field trials over a 0.8-hectare test plot, the reinforcement-learning-guided policy achieved a 92% water-stress-zone detection rate versus 61% for a fixed-grid baseline, while reducing mission flight time by approximately 32%. We further report an ablation study isolating the contribution of SLAM-derived canopy structure to CWSI accuracy, a sensitivity analysis across field complexity, and a full error budget for the fused pipeline. We close with a discussion of validation limitations, broader scientific and agricultural impact, and a roadmap toward multi-UAV fleet deployment for whole-farm monitoring
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 matchEurope PMC · checked 5 Sept 2026
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.
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 matchEurope PMC · checked 15 Sept 2026
Cercospora janseana (Racib.) O. Const. is a re-emerging fungal pathogen that causes Cercospora net blotch on rice. Previous research on resistance to C. janseana has primarily focused on foliar symptoms. Subsequently, sheath infection remains poorly characterized which hinders disease management efforts. This study developed and validated a reproducible sheath inoculation protocol under controlled conditions. Three inoculation methods (agar disc, spray, and drop) were evaluated with and without mechanical wounding. Lesions only formed with inoculation methods using wounding and the agar disc method produced the most consistent and uniform symptom development. Time-course analysis in the susceptible variety Cheniere revealed earlier lesion onset, more rapid expansion, and lower variability in the agar disc method compared to spray, confirming its suitability for phenotypic screening. The optimized protocol was applied across five independent trials involving four rice varieties. DG263L consistently exhibited minimal lesion development, confirming its resistance, while Cheniere showed extensive lesion growth, indicating high susceptibility. PVL03 and LaGrue displayed moderately susceptible reactions, with PVL03 developing significantly higher lesion lengths and AUDPC values than LaGrue in one-month-old plants. Although lesion onset was delayed in 45-day-old plants, disease progressed more rapidly once established. AUDPC analysis corroborated these trends, further distinguishing varietal responses. The protocol effectively discerned resistant, intermediate, and susceptible phenotypes, supporting its use in resistance screening. To our knowledge, this is the first controlled sheath inoculation method developed for Cercospora net blotch, offering a standardized approach for evaluating sheath-specific resistance and advancing the characterization of the C. janseana -rice pathosystem.
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.
The edible yield and commercial value of durian are strongly determined by the number of fully developed internal locules, yet their assessment still relies largely on subjective manual inspection or costly destructive analysis. Existing non-destructive approaches, including tapping-based evaluation and X-ray imaging, are either insufficiently standardized or impractical for high-throughput phenotyping. This study proposes the Rotation-Synchronized Structural Locule Evaluator (RS-SLE). The framework integrates convolutional spatial learning with deterministic signal processing to estimate internal locule number from rotational RGB imaging. The proposed system learns frame-wise spatial morphology using convolutional heatmap regression and performs temporal inference through deterministic kinematic signal processing over a mechanically bounded 360° rotation, thereby avoiding reliance on data-intensive recurrent video models. Specifically, sequential 2D locule-probability heatmaps are transformed into a synchronized 1D morphological energy signal, which is subsequently refined using gradient-based integration, Tikhonov regularization, and adaptive morphological thresholding to recover cycle-consistent structural peaks corresponding to fertile locules. This design provides a transparent and computationally efficient alternative to learned temporal memory while maintaining robustness to viewpoint variation, partial occlusion, and high-frequency spine noise. Evaluated on 260 fruits of the ‘Monthong’ cultivar, the proposed framework achieved a mean absolute error of 0.289 locules and a 71.1% exact-match rate on independent rotational videos. When combined with morphological correlation analysis, the framework attained a 100% tolerance accuracy (±1 locule) on the test set with destructive ground-truth measurements. These results demonstrate that dynamic external morphology can serve as a reliable optical proxy for internal locule development. Crucially, the final locule count is derived entirely from the CNN-generated one-dimensional morphological signal. This demonstrates the necessity of localized structural analysis over simple macroscopic shape indices. The findings highlight the value of explainable, kinematics-informed vision systems for practical precision agriculture.
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.
Red crown rot of soybean (RCR), caused by Calonectria ilicicola, is an emerging soilborne disease whose quantification is challenging due to its complex symptom development across root and foliage levels. This study developed and evaluated a multi-scale framework to improve the assessment of RCR severity from controlled environments to field conditions using root imaging and standardized visual scales. Under controlled conditions, a standard area diagram (SAD) for root necrosis was developed and validated, and SAD-assisted evaluations significantly improved accuracy, precision, and inter-rater agreement compared with unaided assessments. In field conditions, a diagrammatic symptom scale (DSS) was developed using consensus-rated images from experts and showed high reliability, repeatability, and reproducibility across 18 raters, with strong intra- and inter-rater agreement. This study developed and evaluated complementary methods to improve the assessment of RCR severity from controlled environments to field conditions using root imaging and standardized visual scales.
Fenech M, Fernandez-Moreno J, Daubermann G, Nawar A, Taylor J, Davis H, Belcapo S, Budnick A, Yaschenko A, Xu C, Hand H, Jackson J, Vollen K, Muller K, Kater M, Moura D, Ascencio-Ibáñez J, Alonso J, Stepanova A.
ArabidopsisCell / cellular structureRootPhysiological trait estimationGrowth / development / phenology
Decoding how plants integrate multiple hormone signals to coordinate growth requires tools capable of resolving pathway interactions at cellular resolution in living tissue. Here we present ACE (Auxin–Cytokinin–Ethylene) and ACE2 , proof-of-concept single-locus reporters to simultaneously capture activity of multiple hormones. Deploying ACE alongside well-established reporters, exogenous hormone treatments, and reverse-genetic perturbations of hormone biosynthesis, signaling, and transport in three-day-old etiolated Arabidopsis seedlings, we dissect the spatiotemporal hierarchy governing primary root elongation and root apical meristem (RAM) size. We demonstrate that both ethylene- and cytokinin-triggered root growth inhibition involve a boost of TRYPTOPHAN AMINOTRANSFERASE OF ARABIDOPSIS1 (TAA1)-mediated auxin biosynthesis and AUXIN RESISTANT1 (AUX1)-dependent auxin redistribution. Two spatially distinct auxin responses underlie the respective root growth effects: ethylene expands TAA1-dependent auxin biosynthesis from the root vasculature into the epidermis and promotes AUX1-mediated auxin import into the transition and elongation zones to inhibit cell elongation, while cytokinin confines ethylene-dependent TAA1-boosted activity to the vasculature and drives auxin accumulation in lateral root cap cells to reduce RAM size. Together, these data establish a reciprocal regulatory loop between these hormones, positioning ethylene as a convergence node in auxin–cytokinin crosstalk, and cytokinin as a modulator of the ethylene–auxin interaction. Critically, the changes in cross-activated reporter patterns described for different genetic backgrounds, alongside quantitative assessment of hormone-specific inhibition of the mutants’ growth, were consistent with the multi-hormone network established over two decades of research, and added cell-type-resolved spatial detail and a proposed hierarchy for the etiolated seedling root. Finally, a second-generation reporter, ACE2 , overcomes key technical limitations of ACE , expanding the platform’s capacity toward a higher-order multi-hormone monitoring system. These resources expand the Arabidopsis genetic toolkit and provide a generalizable framework instrumental for dissecting multi-hormone signaling hierarchies at the cellular level.
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.
Abstract Probe-based spatial transcriptomics platforms use predefined oligonucleotide panels to detect selected RNAs in tissue sections while preserving transcript spatial coordinates. Accurate cell segmentation is required for reliable transcript-to-cell assignments. This analytical process is affected in plant tissues by cell walls, large vacuoles, and strong autofluorescence, which often reduce boundary contrast and elevate background. Nucleus-only segmentation with fixed-distance expansion can be an alternative approach, but it underestimates cellular area and morphology and reduces the number of assignable transcripts per cell. Here, we present a practical workflow for segmentation and downstream processing in plant probe-based spatial transcriptomics. Using the soybean nodule, soybean seed, rice root, and wheat inflorescence, we demonstrate the applicability of our workflow across species, tissues, and technological platforms. In brief, candidate cell masks are generated from available fluorescence signals and then selected and corrected using two napari plugins. Transcript-informed refinement with Baysor is included as an optional step. Upon benchmarking our approach using a collection of metrics (assignment yield, background/negative controls, and per-cell transcript/gene distributions) and linked segmentation choices to expression-matrix quality and downstream clustering, we demonstrate the potential of our workflow to support the analysis of plant probe-based spatial transcriptomics.
Drought stress severely limits foxtail millet yield and quality, yet current drought-resistance indices are exclusively yield-oriented and ignore grain-filling quality. Our two-year (2024–2025) experiments with 24–48 varieties revealed that yield and blighted grain rate (BGR) are partially decoupled (e.g., Zhangzagu 18: yield 2307 kg/ha, BGR 0.444; Zhonggu 19: yield 1622 kg/ha, BGR 0.280). We therefore constructed the Yield–Quality Synergy Index (YQSI = DYI − BGR), which penalizes varieties with poor grain filling. The YQSI tied for first place with DYI in comprehensive screening performance and achieved the highest inter-annual stability (Spearman ρ = 0.823, Jaccard = 0.438, composite score = 1.261). Sensitivity analysis confirmed robustness of the equal-weight formula across a 4-fold range of quality-penalty weights. Six strongly drought-resistant germplasms with balanced yield and quality were identified. Using UAV multimodal data (RGB, multispectral, and thermal infrared) acquired during grain filling, a Random Forest model predicted a YQSI with overall R2 = 0.819 and an F1 score of 0.933 for variety screening. Feature-importance analysis highlighted NDVI, WDRVI, and red-edge texture as key predictors. This study provides a quality-constrained drought-resistance evaluation framework and demonstrates the potential of UAV-based high-throughput phenotyping for foxtail millet breeding.
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 matchCrossref · checked 5 Sept 2026
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.
Continuous X-ray computed tomography (XCT) combined with digital volume correlation (DVC) is presented to quantify internal three-dimensional strain and failure dynamics in apple cortex tissue during compression, revealing how the cellular microstructure governs its mechanical response. We introduce a dimensionless number Mi that is a function of the average interfacial contact area of cells, cell wall thickness, the average cell volume and tissue porosity, to describe tissue microstructure over different development stages. Mechanical softening during maturation aligned strongly with decreasing Mi, linking microstructure to effective Young's modulus, peak stress, and toughness. DVC revealed distinctive strain-distribution signatures: in young, low-porosity tissue, strain was initially diffuse with early-onset localization indicating progressive failure, whereas mature, high-porosity tissue exhibited sharply peaked strain distributions and highly localized fracture planes indicative of brittle collapse. These findings demonstrate how pore evolution, anisotropy, and cell packing jointly determine macroscopic deformation, establishing XCT-DVC as a powerful framework for connecting plant tissue architecture to mechanical function.
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
Stomatal traits are key microscopic phenotypes for evaluating plant physiology, stress responses, and crop breeding potential. However, in vivo high-magnification microscopy often suffers from a shallow depth of field, causing noticeable defocus blur across different spatial locations and making it difficult to capture clear and complete stomatal structures in a single image. Multi-focus image fusion offers a practical solution, yet existing methods typically rely on supervised training, paired data, or hand-crafted rules, limiting their use in real agricultural microscopy scenarios. In this study, we propose an unsupervised multi-focus fusion framework for reconstructing fully focused stomatal microscopic images. The method integrates two-dimensional feature extraction with three-dimensional cross-focal-plane modeling to capture both spatial details and complementary information across focal planes. A max-response-guided spatial gating module is introduced to enhance focused regions while suppressing defocused responses. Additionally, dual sharpness priors based on perceptual features and wavelet high-frequency information enable pixel-wise pseudo-supervised learning without requiring all-in-focus ground-truth images. The model also predicts a probabilistic focal-plane volume for interpretable all-in-focus reconstruction. Experiments on a maize multi-focus image dataset demonstrate that the proposed method achieves superior or competitive performance across multiple fusion metrics, with entropy (EN), edge information preservation ( Q AB∕F ), Chen-Blum contrast metric ( Q CB ), and visual information fidelity for fusion (VIFF) reaching 7.43, 0.21, 0.41, and 1.01, respectively. Ablation studies confirm the effectiveness of the 3D modeling, spatial gating, and dual-prior sharpness supervision. More importantly, when the fused images serve as input to a YOLO-based stomatal instance segmentation model, the proposed method yields the best segmentation accuracy, with mAP50 and mAP50-95 reaching 0.9937 and 0.9121, respectively. Phenotypic measurements derived from the segmentation masks show high consistency with manual annotations, with the highest coefficient of determination R 2 = 0.97 achieved for stomatal count. These results indicate that the framework can act as an effective front-end module for automated microscopic stomatal phenotyping in agriculture.
Reproduction assets foundThe authors state their data and code are publicly available on GitHub, covering the multi-focus stomatal microscopy dataset and the UMF-stomata fusion/phenotyping code.Code · publicOur data and code are available at: https://github.com/Longer-S/UMF-Stomata.Open asset ↗Longer-S/UMF-Stomatahtml-lines:640-655Dataset · publicOur data and code are available at: https://github.com/Longer-S/UMF-Stomata.Open asset ↗Longer-S/UMF-Stomatahtml-lines:683-756Code / dataset availability confirmedEurope PMC · checked 5 Sept 2026
Etukuri SP, Courtney CL, Rhoden C, Laufer M, Main R, Smith RA, Kenny D, Patel K, Johnson J, Thomas HB, Kothari N, Shumate B, Martin VB, Rife TW, Saski CA.
Abstract Cross-sectional analysis is considered the reference method for measuring cotton fiber fineness and maturity; however, its widespread use has been limited by labor-intensive sample preparation and manual image analysis. The objective of this study was to develop and validate a reproducible deep-learning-based workflow for automated segmentation and quantitative analysis of cotton fiber cross-sections. A total of 249 composite light microscopy images of cotton fiber cross-sections were collected and manually annotated to generate training and validation datasets. A YOLO11m instance segmentation model was developed to identify cotton fiber and lumen regions and automatically extract quantitative traits, including fiber area, lumen area, fiber perimeter, and lumen perimeter. The workflow integrates automated image segmentation, post-processing, quantitative trait extraction, and data export to facilitate reproducible cotton fiber phenotyping. Model performance was evaluated using mean Average Precision (mAP), and workflow outputs were validated against Adobe Photoshop using descriptive comparisons of six cross-sectional traits. The model achieved Box mAP50 scores of 0.984 for cotton fiber regions and 0.789 for lumen regions, demonstrating high segmentation accuracy. The automated workflow substantially reduced manual analysis time while producing measurements with central tendencies comparable to those obtained using Adobe Photoshop. To facilitate reproducibility and adoption, the workflow, trained model weights, and supporting documentation are publicly available through GitHub and a Hugging Face web application. The workflow substantially increases analytical throughput while providing a reproducible and publicly accessible method for automated cotton fiber cross-sectional phenotyping, facilitating quantitative analysis for cotton genetics and breeding research.
Reproduction assets foundThe paper's cotton fiber cross-section phenotyping workflow (YOLO11m segmentation pipeline, trained model weights, example images/outputs) is explicitly stated as publicly available via a GitHub repository and a Hugging Face web application, with URLs matching the allowed list.Code · publicThe complete source code, training scripts, dataset configuration, example input images, example outputs,
and supporting documentation are publicly available through the GitHub repository:
https://github.com/RifeLab/cotton-lumen-microOpen asset ↗RifeLab/cotton-lumen-micropdf-page:11 lines:1-47Code · publicThe cotton fiber image analysis workflow is publicly available through a web-based application hosted on
Hugging Face at: https://huggingface.co/spaces/chaneylc/cotton_fiber_microscopy_measureOpen asset ↗pdf-page:11 lines:1-47Plant phenotyping relevance matchEurope PMC · checked 15 Sept 2026
Guo Y, Ren S, Yu Q, Bai X, Ji S, Chen J, Zhu Y, Xiong J, Fang L, Liu C, Li J, Li Y, Wang Q.
Field / plotMultispectral / hyperspectralClassificationGrowth / time-series analysisGrowth / development / phenology
Aquatic plants are vital for lake ecosystem functioning and water-quality stability, yet their community dynamics and phenology rhythms remain insufficiently understood, due to the lack of effective strategies for fine-scale species mapping and phenology extraction. In this study, based on Sentinel-2 MSI imagery, we developed an integrated framework combining machine learning and a priori ecological knowledge to quantify the spatiotemporal changes in aquatic plant distribution, species composition and phenological dynamics for eight dominant species in five regulating lakes along the Eastern Route of the South-to-North Water Diversion Project in China. Results showed that the proposed framework enabled accurate aquatic plant identification, achieving an overall classification accuracy of 96.16% and over 90% accuracy for each species. Since 2016, aquatic vegetation coverage has substantially declined in most lakes, mainly due to the retreat of submerged vegetation. Community structure has shifted from submerged-plant dominance to emergent and floating-leaved dominance in two of them. Phenologically, we found that most aquatic vegetation exhibited a longer growing season, characterized by earlier growth onset (-0.28 days/year) and peak timing (-0.67 days/year) and delayed senescence (0.54 days/year). Correlation analysis indicated that aquatic vegetation dynamics was associated with climate variation, nutrient enrichment, turbidity, and water diversion, with warming and solar radiation likely promoting the growth of some emergent species, while nutrient enrichment and turbidity could be linked with submerged vegetation decline and earlier phenological shifts. Overall, this study provides an effective framework for species-level mapping and phenological monitoring of aquatic vegetation, offering valuable support for the management and conservation of lake ecosystems.
Efficient facility-scale tomato ripeness monitoring remains difficult in greenhouses where uneven terrain limits conventional wheeled and rail-guided platforms and planar cameras provide restricted coverage. This study developed a wheel-legged quadruped monitoring system integrating LiDAR, a depth camera, and a panoramic camera. An adaptive gait-switching strategy supported navigation across heterogeneous terrain. Panoramic images were projected into six perspective views, and the left and right views were processed using a YOLOv8-based ripeness recognition model. Time-synchronized detections and robot poses were fused to map ripeness observations into three-dimensional greenhouse coordinates. Five field experiments in a commercial tomato facility demonstrated autonomous row traversal, inter-row transition, and avoidance of pedestrians, obstacles, and cultivation boundaries. The recognition pipeline continuously identified multiple ripeness stages under variable illumination, foliage occlusion, and robot motion, while the spatial fusion procedure produced a facility-scale three-dimensional ripeness distribution. The integration of terrain-adaptive quadruped mobility, panoramic perception, and spatial mapping provides a practical framework for continuous ripeness monitoring and can support targeted harvesting, yield forecasting, and crop management.
Abstract Plants employ non-photochemical quenching (NPQ) to protect their photosynthetic apparatus from photodamage. The response latency of NPQ following changes in light intensity is thought to significantly decrease photosynthetic efficiency. The amount of NPQ is commonly quantified from chlorophyll-fluorescence techniques using the Stern–Volmer equation, which requires fully closed reaction centres (RCs) of photosystem II, yielding NPQ in the absence of photochemical quenching ( $${\rm{NPQ}}^{\rm{Closed}}$$ NPQ Closed ). However, in nature, NPQ and photochemical quenching are normally present simultaneously. Therefore, to obtain a full understanding of this process, NPQ should also be explored when the RCs are open. Here we developed two methodologies to obtain NPQ in the presence of photochemistry ( $${\rm{NPQ}}^{\rm{Open}}$$ NPQ Open ) using both fluorescence lifetime and fluorescence yield measurements. A detailed comparison in Arabidopsis thaliana plants reveals that the value of $${\mathrm{NPQ}}^{\mathrm{Open}}$$ NPQ Open is ~35% lower than that of $${\rm{NPQ}}^{\rm{Closed}}$$ NPQ Closed . This difference is consistently observed across all measurements and is seen both upon closing ( $${\rm{NPQ}}^{\rm{Open}}\to {\rm{NPQ}}^{\rm{Closed}}$$ NPQ Open → NPQ Closed ) and upon reopening ( $${\mathrm{NPQ}}^{\mathrm{Closed}}\to {\mathrm{NPQ}}^{\mathrm{Open}}$$ NPQ Closed → NPQ Open ) of the RCs. We show that this difference can be explained by the presence of RC-induced ‘instantaneous’ switching of the NPQ quenching rate. This means that, in plants, NPQ is much more economical than is widely believed, it is large when its presence is needed, and it decreases instantaneously when the need disappears.
Reproduction assets foundThe paper's custom ultrafast fluorescence analysis code (ICA-based PSII/PSI deconvolution and NPQ calculations) is explicitly deposited by the authors on GitHub, alongside the original data contributions.Code · publicr(s) for their contribution to the peer review of this work. Peer reviewer reports are available.
Funding
This work was supported by ‘Nanoscale regulators of photosynthesis’ NWO research project (project number: OCENW.GROOT.2019.86).
Data availability
The original contributions presented in the study are available via GitHub at https://github.com/L-Ramakers/Heimdall .
Code availability
The custom analysis code used in the study is available via GitHub at https://github.com/L-Ramakers/Heimdall .
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional afOpen asset ↗L-Ramakers/Heimdalllines:88-125Plant phenotyping relevance matchCrossref · checked 5 Sept 2026
The uniformity of maize plant spacing serves as a critical indicator for assessing sowing quality, seed vigor, and field seedling emergence stability. However, manual measurement is inefficient, and complex field conditions make automatic seedling detection and plant spacing measurement challenging. Aiming at the challenges of missed detection, insufficient accuracy for small targets, and large errors in plant spacing calculation under complex field conditions, this study constructs a high-quality dataset containing 693 maize seedling images and implements preprocessing enhancement for images degraded by haze or dust. An intelligent maize seedling detection and plant spacing measurement method based on improved YOLOv8 is proposed. The Global Attention Mechanism (GAM) is embedded into the backbone network to strengthen cross-dimension information interaction between channels and spaces, suppress background interference, and reduce the missed detection rate. The Bi-directional Feature Pyramid Network (BiFPN) is adopted to replace the original PAFPN for enhanced multi-scale feature fusion and deep semantic representation. A new 160 × 160 high-resolution small-object detection layer is added to significantly improve the detection performance of weak and small seedlings. Experimental results demonstrate that the improved model achieves a precision, recall, mAP50, and mAP50-95 of 89.4%, 90.3%, 94.4%, and 49.4%, respectively, which are 2.6, 0.5, 1.5, and 2.7 percentage points higher than those of the original YOLOv8 model. These results indicate that the proposed model improved maize seedling detection performance under complex field conditions. Automatic plant spacing calculation is realized based on detection outputs; the average plant spacing of the dataset is 30.42 cm, with a relative error of only 4.93% compared with the preset sowing spacing of 32 cm. The proposed method can efficiently accomplish field seedling identification, plant spacing quantification, and sowing quality evaluation, providing reliable technical support for precision maize sowing, seeder parameter optimization, and intelligent field management, which is of great significance for promoting the intelligent upgrading of grain crop production.
Rodríguez-Jiménez DM, Dolinko AE, Galizzi GE, Caputo C.
BarleySeed / grainPhysiological trait estimationGrowth / development / phenologyWater status / transpiration
Background and aims Laser Biospeckle Activity (LBSA), derived from laser-induced speckle variations in response to dynamic changes in living tissues, is a promising non-invasive technique for evaluating seed germination. Methods Malting barley (Hordeum vulgare subsp. distichum L. cv. Sinfonia) seeds were analysed using five coefficients-Generalised Differences (GD), Fujii, Lasca, Frequent Motion Image (FMI), and Moment of Inertia (MI)-to assess their ability to discriminate between treatments and tissue regions, and to track changes during imbibition. Whole and longitudinally cut seeds from two treatments (control [untreated] and autoclaved [heat-inactivated]) were analysed, focusing on embryo/endosperm activity ratios. LBSA was also evaluated as a function of imbibition time and seed moisture content. Key results Four coefficients (GD, Fujii, FMI, and MI) successfully differentiated control and autoclaved seeds, as well as embryo and endosperm regions in control seeds, revealing distinct activity patterns. In control seeds, LBSA increased with imbibition time and was well described by polynomial models (quadratic for Fujii and MI; cubic for GD and FMI). GD, Fujii, and FMI required a minimum seed moisture content of 25% to detect activity, while MI was responsive only above 31.5%. In contrast, Lasca was exclusively sensitive to hydration level, fitting a relaxation curve independent of treatment. Conclusions LBSA constitutes a robust, non-destructive methodology for monitoring early germination processes. By combining coefficients, it is possible to infer physiological traits such as embryo specificity, hydration thresholds, and dynamic metabolic reactivation. This positions LBSA not only as a diagnostic tool for seed viability but also as a physiologically informative proxy for studying germination and tissue-level dynamics.
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:1-55Plant 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.
The inflorescence is a key yield-determining organ, yet its complex morphology makes manual pixel-level annotation time-consuming, leading to a scarcity of high-quality segmentation datasets. To address this bottleneck, we propose ZCAT (Zero-shot Cross-crop Annotation Transfer), a novel paradigm for zero-annotation cross-crop pseudo-mask screening. ZCAT completely eliminates pixel-level manual annotation of the target crop, requiring only holistic quality assessment of model-generated pseudo-masks (5-10 s per image). Specifically, we train a SegFormer model on public rice panicle datasets (CVRP and RiceSEG) and transfer it across crops to the wheat spike segmentation task. The key innovation is the introduction of a human-defined quality function Q, which circumvents the fundamental challenge in self-learning algorithms: the inability of computers to autonomously distinguish good masks from bad ones. Through iterative human-in-the-loop pseudo-label screening with a curriculum learning strategy, each round adds only a few high-quality pseudo-masks to the training set, continuously improving model performance. After four iterations, ZCAT produces pseudo-masks with an average Spike IoU of 0.7003, evaluated against the GWFSS manual annotations as ground truth. Moreover, the pseudo-mask dataset exhibited higher benchmark performance than the GWFSS manual annotations (Spike IoU 0.7612 vs. 0.7027; mIoU 0.8627 vs. 0.8247), suggesting stronger self-consistency. A generalization test on a strictly held-out set of 100 manually annotated wheat spike images showed that the model trained on ZCAT-generated pseudo-masks achieved marginally better performance than that trained on the GWFSS manual annotations (Spike IoU: 0.5112 vs. 0.4927; mIoU: 0.5627 vs. 0.5247). The time budget of the ZCAT pipeline was substantially lower than that of manual annotation. ZCAT opens a new pathway for rapid annotation of plant reproductive structures or other organs and significantly reduces data preparation costs in plant phenomics. The generated wheat spike pseudo-mask dataset and the mask quality screening tool (Mask Quality Screener) are open-sourced.
Reproduction assets foundThe paper explicitly open-sources two paper-specific assets: the ZCAT-generated wheat spike pseudo-mask dataset and the Mask Quality Screener tool, both with public GitHub URLs in the Data availability statement.Dataset · publicThe wheat spike pseudo-mask dataset and Mask Quality Screener are available at https://github.com/zyxyes1/MaskQualityScreener and https://github.com/zyxyes1/Wheat-Spike-Semantic-Segmentation , respectively.Open asset ↗Wheat-Spike-Semantic-Segmentationlines:415-440Plant phenotyping relevance matchOpenAlex · Crossref · checked 5 Sept 2026
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.
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 matchEurope PMC · checked 15 Sept 2026
Understanding the dynamic regulation of endogenous metabolites in plants under environmental stress is essential for elucidating plant adaptive mechanisms and improving crop resilience. However, conventional analytical methods are typically destructive, time-consuming, and lack the capability for real-time monitoring, thereby limiting the investigation of in-vivo biochemical dynamics in plants. In particular, the in-situ, non-invasive detection of small-molecule regulators such as protocatechuic acid (PCA) remains a significant challenge. Herein, we report a wearable electrochemical sensing platform based on a Cu/ZIF-8-modified flexible printed electrode (FPE) for real-time, in-situ monitoring of PCA in plant leaves. The incorporation of Cu into the ZIF-8 framework enhances the electrical conductivity and electrocatalytic activity of the material while maintaining its porous structure, enabling sensitive detection of PCA. By integrating with reverse iontophoresis (RI), non-invasive extraction and continuous monitoring of PCA from living plant tissues are achieved. The dynamic behavior of PCA in green tea plants under light deprivation and drought stress is systematically investigated. The results reveal distinct stress-dependent response patterns, with PCA levels rapidly decreasing under both dark and drought conditions, highlighting its critical role in stress adaptation and metabolic regulation. This work establishes a versatile strategy for real-time tracking of endogenous plant metabolites and provides new insights into plant physiological responses under environmental stress. The proposed platform holds significant promise for applications in plant science, precision agriculture, and the development of stress-resilient crops.
Abstract China's rice production and environmental sustainability are largely dependent on the cold black soil region in Northeast China, where precise water and nitrogen management is challenged by water scarcity and high carbon emissions. To overcome the limitations of conventional empirical management and improve the accuracy of evapotranspiration (ET) estimation in controlled-irrigation paddy fields, this study proposes a novel framework integrating unmanned aerial vehicle (UAV) multispectral and thermal infrared observations, the FAO-56 dual crop coefficient approach, and the NSGA-II multi-objective optimization model. To parameterize and validate this methodology, field data comprising four lower limit thresholds for controlled irrigation and four nitrogen fertilizer application rates were acquired from the Rice Research Site of Farm 856, Heilongjiang Province, China. This integrated approach was used to systematically evaluate rice growth, water consumption, resource use efficiency, and greenhouse gas emissions under different water-nitrogen treatments. Based on these evaluations, an irrigation optimization scheme was developed using daily crop evapotranspiration (ETc). The results indicated that water, nitrogen, and their interaction significantly affected rice yield, irrigation water use efficiency (IWUE), partial factor productivity of nitrogen (PFPN), and global warming potential (GWP). Treatments W3N2 (80%+155 kg/ha N) and W3N3 (80%+200 kg/ha N) achieved the highest yields, 11,883.51 and 11,436.82 kg/ha, respectively, whereas W2N1 (70%+110 kg/ha N) exhibited the best comprehensive performance, with a TCQ value of 0.65. Among the tested vegetation indices, the normalized difference vegetation index (NDVI) showed the strongest correlation with the basal crop coefficient, with an R²of 0.85. The NDVI -crop water stress index ( CWSI ) coupled model achieved the highest ET c estimation accuracy (R 2 = 0.89, RMSE = 0.39 mm/day), reducing the RMSE by 10.3% compared to the traditional, Multi-objective optimization revealed obvious trade-offs among high yield, water saving, high nitrogen efficiency, and low emissions. Scenario S5 was identified as the optimal solution, with an irrigation amount of 669.94 mm, a nitrogen rate of 117.48 kg/ha, a yield of 11,473.43 kg/ha, and the highest coordination degree of 0.86. These results demonstrate that coupling UAV multispectral and thermal infrared imagery with the FAO-56 model can effectively improve ETc estimation and provide reliable data support for water-nitrogen multi-objective optimization in cold-region rice production.
Thermal imaging enables non-invasive assessment of canopy temperature, an essential indicator of plant stress, yet the lack of color cues and strong shadow interference make plant segmentation in thermal images difficult. Recent advances in foundation models have demonstrated improved performance and generalizability across applications, showing promise for domain-specific applications with limited annotated datasets such as plant segmentation in thermal images. This study investigates large multimodal models (LMMs) for thermal image segmentation in plant phenotyping. Building upon the SAM-CLIP framework, we design a unified pipeline spanning zero-shot inference, few-shot and low-shot fine-tuning, and active learning to maximize accuracy with minimal supervision. Evaluations on two thermal datasets, LadyBird Brassica and UGA Brassica, demonstrate robust performance after minimal adaptation across both datasets and superior performance compared with baselines, achieving mIoU D values of 97.54% on the LadyBird dataset and 76.94 % on the UGA dataset. We also release the resulting thermal segmentation annotations to support community benchmarking and reproducible research, highlighting the potential of LMMs to enable scalable, high-quality dataset construction for field phenotyping. The released datasets can be found at: https://cornell.box.com/s/dh69xf84464yrc1vlws92l1tflx7qa89
Reproduction assets foundThe authors publicly released the paper-specific thermal segmentation annotations (20,538 LadyBird masks and 37,790 UGA masks) via a Cornell Box link stated in the abstract, results, and data availability statement. No author analysis code or trained model checkpoints are explicitly released; the mmsegmentation GitHub/Dataset · publicwe generated and publicly released segmentation annotations for the complete LadyBird and UGA thermal image datasets using the best-performing SAM-CLIP model. Specifically, the final model obtained through the multi-round training process was used to generate 20,538 masks for the LadyBird dataset and 37,790 masks for the UGA dataset. Details of the generated annotations are provided in Supplementary Fig. S1 , and both annotated datasets are publicly available at: https://cornell.box.com/s/dh69xf84464yrc1vlws92l1tflx7qa89Open asset ↗lines:220-232Plant phenotyping relevance matchCrossref · checked 14 Sept 2026
Published12 Aug 2026International Journal of Computer Information Systems and Industrial Management ApplicationsCited by 0 · OpenAlex ↗
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 matchCrossref · checked 15 Sept 2026
India’s economy is heavily reliant on agriculture, with a diverse array of crops grown on vast tracts of land. Fruit cultivation, especially papaya, has become more popular in recent years because of its high nutritional and financial value. To increase yield, maximize resource use and lessen reliance on chemical pesticides, modern techniques like protected cultivation and hydroponics are being used more and more. Fruit crops grown in controlled or semi-controlled environments are still susceptible to nutrient imbalances despite these developments, which can have a substantial impact on plant health, fruit quality and overall productivity. Papaya leaf nutrient deficiencies frequently show up in the early stages of growth and can result in poor fruit development and decreased yield if they are not detected in time. To support early diagnosis and better crop management, the current study focuses on creating an effective method for identifying nutrient deficiencies in papaya leaves using a deep learning (DL) framework based on transfer learning (TL). In this nutrient and micronutrient deficiency study and field work observation during year 2024 to 2026 with different climate and weather conditions done in order to tackle a new but related classification task, in the context of plant health assessment, several well-established architectures including InceptionV3, VGG19, DenseNet and Xception have been widely explored for leaf image analysis. Studies commonly utilize publicly available datasets, such as papaya leaf image repositories hosted on platforms like IEEE DataPort, to fine-tune these models for efficient feature extraction and accurate identification of nutrient and micronutrient deficiency patterns. This body of work demonstrates the growing role of transferring convolutional neural network (CNN) models in advancing automated crop monitoring and decision support systems.
Accurate estimation of the photorespiratory CO 2 compensation point (Γ*) is essential for describing the balance between Rubisco carboxylation and oxygenation and for parameterising biochemical models of photosynthesis. Γ* and the rate of CO 2 release in the light (D L ) are commonly estimated using the Laisk method, based on measurements of net CO 2 assimilation rate (A net ) at low chloroplastic CO 2 concentrations (c c ), under several sub-saturating irradiance levels. However, many widely used temperature dependence relationships for Γ* (Γ*(T)) were derived using conventional linear implementations of the Laisk method, despite the intrinsically nonlinear behaviour of the A net -c c response predicted by the photosynthetic theory. Here, we revisited the temperature dependence of Γ* and D L using the improved Laisk-FvCB framework that simultaneously constrains the nonlinear A net -c c response across multiple irradiance levels. Gas exchange of sunflower leaves was measured across a wide temperature range from 3.9°C to 42.0°C. The conventional linear implementation generated highly dispersed pairwise intersections and unstable estimates of both Γ* and D L , including some physiologically unrealistic negative D L values at low temperatures. In contrast, the mechanistically constrained Laisk-FvCB framework produced physiologically meaningful temperature responses and substantially reduced methodological artefacts associated with linear extrapolation. Using this framework, we derived a revised in vivo Γ*(T) relationship described by an Arrhenius-type function with Γ*(25) = 43.4 μmol mol -1 and an apparent activation energy of 27.7 kJ mol -1 , such that Γ*(T) = 43.4 exp[11.176 ((T - 25)/(T + 273.15))], where T is leaf temperature in °C. Comparison with other widely used Γ*(T) formulations showed substantial divergence at temperature extremes, often exceeding the variability expected from realistic interspecific differences in Rubisco specificity among C 3 species.
Yang H, Zeng L, Gou X, Shao Y, Song X, Zhang S, Qian X, Zhang Z, Wang X, Li H, Peng B, Anyanwu JN, Zheng Z, Han L, Zhou L, Bulut M, Song P, Yang W, Xiao Y, Li W, Dai M, Qiu F, Zhang J, Wang B, Fernie AR, Wang H, Zhao H, Li X, Yan J, Guo T.
Improving nitrogen use efficiency (NUE) is essential for sustainable agriculture, yet conventionally measured plant characteristics have limited value as NUE proxies. Here we show that artificial intelligence (AI) can uncover previously unrecognized phenotypic variation associated with NUE, revealing genetic variation that is largely missed by conventional phenotypes. We trained a convolutional neural network (CNN) on 25,080 maize images to learn features that distinguish how plants respond to low- and high-N conditions, achieving 96.7% accuracy. The learned features were defined as deep phenotypes. Compared with conventional phenotypes, deep phenotypes showed greater phenotypic variation and higher heritability, enabling the identification of 523 significant loci compared with 21 for conventional phenotypes. We next investigated candidate genes underlying these loci and used these findings to interpret the learned features. Lower CNN layers primarily reflected visual patterns overlapping with conventional phenotypes, whereas deeper layers encoded additional features associated with N-responsive genetic variation. To validate candidate genes identified by the AI framework, we functionally characterized Liguleless2 (LG2), a basic-leucine zipper (bZIP) transcription factor, and demonstrated that lg2 mutants exhibit enhanced root architecture and increased N uptake efficiency. Field trials of 200 hybrids across diverse N environments further supported the AI findings, with each beneficial allele increasing ear weight by an average of 18 g per plot under low-N conditions. These results show how integrating AI and biology can uncover biologically relevant variation underlying complex traits such as NUE and enhance the interpretability of AI models.
Conventional methods of plant distance and volume measurements are limited by low efficiency, limited spatial coverage, and high measurement error. LiDAR and RGB-D imaging offer cost-effective, precise, and non-destructive techniques for plant distance and volume measurements. This study aimed to measure cabbage height, volume, and distance using LiDAR and RGB-D imaging. The sensors were mounted on a 1.6 kW electric field scouting platform (EFSP) for data collection. Point cloud (PCD) data were collected using LiDAR, whereas data processing, visualization, and measurements were done using commercial software and open-source programming scripts. A total of 20 cabbage plants were analyzed. LiDAR data processing included data frame screening, outlier removal, denoising, voxelization, and generation of 3D PCD density maps. Depth image processing included importing raw data and metadata shaping using intrinsic camera parameters, visualization, extraction of depth points, and pixel-level measurements of distances and volume. RGB image processing involved image conversion, segmentation, normalization, binary masking, mask cleaning, region extraction of cabbages, separation of ROI and preparation of contours, Delaunay triangulation and convex hull preparation, ROI overlay, bounding box preparation, sharing boundary between two boxes, conversion to pixel distances, and for visualization, plant height, volume measurements, and center to center distance measurement for measuring the plant distance. LiDAR demonstrated higher measurement accuracy for cabbage plant height, circumferential volume (geometric canopy volume), and plant distance, followed by RGB-D imaging, while RGB imagery showed comparatively lower performance under the study field conditions. Overall, LiDAR and RGB-D imaging provided reliable and non-destructive approaches for cabbage geometric characterization under field conditions, although accurately capturing complex plant geometry remains challenging. Positive and negative values of bias represent the over- and under-estimated results, respectively. Future studies should include larger and more diverse plant datasets exhibiting diversified size, shape, and geometric structure to further improve the robustness and general applicability of the proposed sensing approaches.
White lupin ( Lupinus albus L.) is a cool-season grain legume with seed crude protein of 33-47%, competitive with soybean ( Glycine max L.) meal. It also fixes nitrogen and mobilizes soil phosphorus. Because soybean is a summer crop, white lupin can occupy Southeastern winter fields as a complementary protein source. Breeding for seed protein is limited by the cost and throughput of reference phenotyping. To determine how each is best deployed, we compared the utility of near-infrared spectroscopy (NIRS)-based phenomic selection with genomic selection based on 246,847 SNPs from low-pass, whole genome sequencing in a panel of Auburn University breeding lines and USDA National Plant Germplasm System germplasm. A handheld NIR calibration against Dumas reference protein reached screening-grade accuracy (R 2 = 0.81). Under common cross-validation, phenomic predictive ability was 0.93 and genomic was 0.12. The low genomic value was consistent with moderate heritability (H 2 = 0.33) and strong genotype-by-year interaction. Beyond predictive ability, NIRS recovered superior accessions the strictest selection intensity, and 40 to 60 reference assays sufficed to calibrate the model. Handheld NIRS is a low-cost tool for protein calibration and early-generation screening, while genomic prediction remains suited to parental selection, together supporting a complementary strategy for legume breeding Plain Language Summary Soybean meal is the main protein source for livestock and fish farms in the United States. Because soybean is a summer crop, many Southeastern fields sit idle or grow low-value cover crops in winter. White lupin, a cool-season legume whose seeds are as protein-rich as soybean meal, makes a good complementary winter crop: it yields high-protein grain while serving as a cover crop that fixes nitrogen and frees up soil phosphorus for later crops. In our early-stage lupin breeding program, measuring seed protein by standard lab methods is slow and costly. We built a calibration that lets a handheld scanner estimate protein from light, and compared it with predicting protein from the plant’s DNA. The scanner gave accurate, low-cost protein screening from only about 40-60 lab tests, while DNA-based prediction remains suited to guiding parent selection. Used together, these tools offer breeders a practical path to develop high-protein white lupin. Core ideas Handheld NIRS provides screening-grade prediction of white lupin seed crude protein. Spectra carried more usable protein signal than markers by measuring seed chemistry directly. NIRS and genomic prediction serve different stages of a white lupin breeding program. About 40 to 60 reference assays sufficed to calibrate NIRS to near-full accuracy.
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.
Gu Y, Sanow S, Taylor T, Morimoto KW, Nemer A, Hadley DJ, Zafar SA, DeMello L, Chen Y, Knab H, Busch Castro A, Kumaravelu V, Bailey-Serres J, Carney R, Brady SM.
Root anatomical barriers, including the suberized and lignified walls of the endodermis and exodermis, and cortical aerenchyma, regulate water and nutrient transport, gas exchange, and rhizosphere interaction. Their adaptive function places them as an important target for breeding environmentally resilient plant species. Quantifying these structures at high resolution is a manual bottleneck that limits experimental scale. We present RADIX (Root Anatomy Deep- learning Image segmentation across species and platforms), a framework that adapts a large self-supervised vision-transformer foundation encoder (DINOv3), pre-trained on billions of natural images, to root anatomy by fine-tuning its encoder with a dense-prediction-transformer decoder. Transferring these general-purpose vision encoders to a specialized biological domain with a high-quality annotated dataset is what allows RADIX to generalize across species and imaging platforms. We train and evaluate it on the first expert-annotated benchmark of root anatomical structures at scale, comprising 1,695 high-quality fluorescence images spanning 17 monocot and dicot species, six anatomical structures, and three imaging platforms. RADIX segments all six structures at inter-annotator-level accuracy and generalizes to unseen species, genotypes, growth conditions, and an imaging platform from an independent laboratory. A single unified model surpasses monocot- and dicot-specialist models without sacrificing in-group accuracy. Predicted masks yield aerenchyma and suberin/lignin measurements matching expert annotation at ∼1.2 s per image with a single GPU, reducing weeks of manual analysis to minutes. Applying RADIX across genotypes, microbial treatments, and growth systems, we show that these cell type features form a coordinated, multidimensional, and context-dependent system shaped by genetic and environmental factors.
Abstract Depth cameras measure the distance to a fruit's surface, whereas converting its silhouette into physical dimensions requires the depth of its center; corrections for this offset have so far been empirical, and therefore bound to the crop, sensor, and dataset they were fitted on. This paper derives the correction analytically. For a spherical fruit, integrating the surface-depth distribution over the visible hemisphere yields a closed-form sphere correction whose coefficient follows from sampling geometry, together with a theoretical justification of the median mask depth as the representative statistic. Combined with deep instance segmentation on RGB-D imagery of hydroponic melons, the empirically optimal coefficient coincided with the derived value, and the pipeline reached R 2 of 0.966 for fruit length (MAE 1.43 mm), 0.959 for width (1.84 mm), and 0.861 for end-to-end fresh weight (MAPE 4.6%). The analytical form made the measurement transferable. Applied unchanged to cylindrical mini-cucumbers, the pipeline held mm-level accuracy (width MAE 0.52 mm; fresh weight R 2 0.955 after coe cient refitting), with the correction's negligibility predicted in advance by an R / Z corollary; across active-stereo and time-of-flight cameras, the optimal coefficients proved non-interchangeable, identifying the coefficient as a physical parameter that absorbs geometry, sensor physics, and fruit shape. A field system that fuses and cross-verifies the two sensors, with an error-propagation confidence gate and parameterized grading logic, reproduced 2-3% fresh-weight error and 90.9% confirmed-judgment grading accuracy over four validation sessions in a commercial greenhouse unseen during development. Throughout, geometric components transferred unchanged while learned and regression components required recalibration - a boundary the model predicts and the system itself monitors.
Common beanLeafClassificationDisease symptoms / severity
Abstract Accurate recognition of plant leaf diseases from images is essential for intelligent agriculture and precision crop protection. However, reliable disease identification remains challenging because lesion regions often exhibit subtle visual differences, complex backgrounds, and large intraclass variations, especially when available disease samples are limited. To address these challenges, this study proposes PR-CNN, a deep learning framework that integrates convolutional neural networks, pyramid split attention, and a relation network for bean leaf disease image recognition. The convolutional backbone is first used to extract visual features from support and query images. Then, the pyramid split attention module enhances multiscale spatial and channel feature representation, enabling the model to focus on discriminative lesion regions while suppressing redundant background information. Finally, the relation network learns a nonlinear similarity metric between paired samples and generates relation scores for disease category prediction. Experimental results show that PR-CNN achieves an overall classification accuracy of 99.24% on the primary bean leaf disease dataset, outperforming representative models, including ResNet50, DenseNet, Inception v4, and EfficientNet B7, in terms of recognition accuracy and adaptability. In addition, PR-CNN was evaluated on four publicly available plant disease datasets, including CGIAR, Plant Diseases, LWDCD 2020, and Plant Pathology, achieving an average accuracy of 99.84%. These results demonstrate that PR-CNN can effectively improve image based plant disease recognition and provides a robust visual classification framework for intelligent crop disease diagnosis.
Grass seed crops are susceptible to yellow dwarf viruses transmitted by aphids. The Willamette Valley in Oregon, United States, is the leading producer of cool-season grass seed crops globally, and industry reports have attributed seed yield loss and shortened stand longevity to aphid-transmitted yellow dwarf viruses. Genetic resources are needed for effective and sustainable management of this pest, specifically the Rhopalosiphum padi–PAV pathosystem, in grass seed production to reduce foliar insecticide applications and maintain optimum seed yield potential. High-throughput phenotyping methods are needed to screen grass seed cultivars to identify resistant traits for traditional breeding programs. An automated video tracking procedure was optimized to evaluate host plant resistance in cool-season grass seed crops to R. padi–PAV with live plants and viruliferous and nonviruliferous aphid populations. Feeding behavior recorded with automated video tracking was strongly correlated with “ground-truthed” observations by human observers. Partial resistance (antixenosis and antibiosis) and tolerance traits were detected in select perennial ryegrass and tall fescue cultivars evaluated with traditional phenotyping methods in a greenhouse setting and with high-throughput phenotyping using automated video tracking in the laboratory. Across grass cultivars, nonviruliferous aphids had greater fitness and preference for noninfected grass plants compared with viruliferous aphids. Automated video tracking can be used as a high-throughput phenotyping method for continued evaluation of host plant resistance in grasses grown for seed production, as well as to identify resistant genotypes in other grass crops susceptible to aphid–yellow dwarf virus virus–vector systems.
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.
Citrus Huanglongbing (HLB) is one of the most destructive citrus diseases worldwide, and early diagnosis remains challenging because uneven pathogen distribution often leads to false-negative PCR results. Here, a paper-based nanozyme sensor array was developed for volatile organic compound (VOC) chemical fingerprinting of HLB-associated citrus samples. Fe/Al bimetallic NH₂-MIL-53 nanozymes were designed to regulate VOC adsorption and peroxidase-like catalytic activity. Exposure of the nanozyme sensors to VOCs reduced the catalytic oxidation of 3,3',5,5'-tetramethylbenzidine (TMB), generating concentration-dependent colorimetric responses. By tuning the Fe/Al ratio, the nanozymes exhibited differentiated responses toward HLB-associated volatiles, including methyl salicylate, phenylacetaldehyde, and linalool, with recognition limits of 0.1-0.5 ppm. Integration with two MOF-based sensing channels formed a five-channel artificial olfactory array capable of generating multidimensional color fingerprints. The sensor array successfully discriminated healthy, asymptomatic, and infected citrus leaf samples, achieving an overall classification accuracy of 92.5%. These results suggest that the proposed platform provides a simple and low-cost approach for citrus VOC chemical fingerprinting and may provide useful information for quality-related screening in citrus production systems, while further field-oriented validation is still needed to assess its practical applicability.
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 matchEurope PMC · checked 15 Sept 2026
Published10 Aug 2026International journal of biological macromoleculesCited by 0 · OpenAlex ↗
Starch, a key biological macromolecule accounting for 50-80% of dry weight in sweetpotato (Ipomoea batatas [L.] Lam.) storage roots, underpins food and industrial applications. However, sweetpotato starch characterization is limited by local-sectioning approaches that fail to capture the whole-root granule dynamics. Here, we established a new morphological observation system covering three key root regions based on two representative cultivars: Okinawa 100 (V100), and Yanshu25 (Y25). It was effective and convenient for in situ starch observation and analysis in sweetpotato roots. The whole-root in situ microscopy, starch physicochemical profiling, and transcriptomic correlation were integrated to resolve starch dynamics in Y25 and V100. We identified widespread simple starch granules (SSGs)-compound starch granule (CSG) coexistence across the whole root tissues, with Y25 exhibiting programmed CSG fragmentation driven by ARCs/FtsZ-mediated amyloplast envelope destabilization and concomitant AMY/BMY upregulation. Y25 had a higher amylose content and a higher proportion of medium/long chains, but the average degree of polymerization was slightly lower. Transcriptomic analyses revealed that the differentially expressed genes were annotated in pathways of carbohydrate metabolism, and the differentially expressed genes in the starch metabolism pathway were analyzed. Weighted gene co-expression network analysis further identified the hub genes from different modules and analyzed the co-expression networks. This work will not only advance the understanding of starch granule assembly and remodeling in sweetpotato, but also provide a robust methodological and transcriptome-guided framework for starch-focused germplasm screening and quality improvement.
Herein, we rationally designed and synthesized two types of D-π-A red-emitting fluorescent probes with intrinsic intramolecular charge transfer (ICT) and twisted intramolecular charge transfer (TICT) characteristics, for traditional methods for lipid droplet detection in food crops suffer from reliance on bulky instruments, tedious sample pretreatment and the impossibility of on-site quantification. To modulate the molecular skeleton, methoxy groups were introduced, and highly specific recognition of lipid droplet microenvironments was realized through the probes' synergistic spectral response to polarity and viscosity. Typical turn-on fluorescence responses and favorable lipophilicity were displayed by both probes, thus enabling accurate localization of the hydrophobic domains of lipid droplets. Possessing ratiometric fluorescence performance, the modified probe OCH 3 -Phe-FCN efficiently alleviated signal interference from complex matrices, and toward oleic acid was a lower limit of detection presented. By combining smartphone-based RGB colorimetry and image grayscale analysis was a portable detection platform constructed, with fluorescence signals employed to accomplish visual analysis of lipid droplets in food crops. Satisfactory biocompatibility and targeting capability were validated for the as-prepared probes via cell imaging and cytotoxicity evaluations, and far superior imaging quality and signal-to-noise ratio were demonstrated by the modified probe. With the aid of density functional theory (DFT) calculations, systematically elucidated was the spectral mechanism of excited-state transitions synergistically modulated by the two microenvironmental factors. Not only does this work deepen the insight into the response mechanism of dual-responsive fluorescent probes, but also a promising technical strategy is offered for in situ spectral detection of lipid droplets in food crops.
LiDAR (Light Detection and Ranging) and RGB-D camera imaging have emerged as essential tools in agricultural applications, particularly for plant size and distance measurements, enabling non-destructive, cost-effective, and precise estimation. The objective of this study was to measure the plant canopy dimensions and distance between apples using commercial LiDAR, and an RGB-D camera with a speed sprayer platform was used to determine whether LiDAR provides a higher measurement accuracy under field conditions. Data were collected in an apple orchard in Muju, Republic of Korea. Commercial 3D LiDAR, a terminal box, an RGB-D camera, a microcontroller, a power supply, and individual display monitors were integrated into a customized data acquisition (DAQ) box for LiDAR point cloud (PCD), RGB, and depth imagery data collection. Commercial software was used for data acquisition, data conversion (pcap to PCD), segmentation of regions of interest (ROI), and pre-processing of data. PCD processing and measurement consisted of data frame selection, data conversion, outlier removal, downsampling, denoising, ground point removal by filtering, voxelization, and density map generation using an open access programming language script. Depth image processing included importing raw data, shaping metadata using intrinsic camera parameters, visualizing depth images, extracting depth points, and measuring the plant canopy at the pixel level. RGB image analysis involved grayscale conversion, thresholding, segmentation of ROI, contour preparation, noise removal, and binary masking for eliminating the background. Estimated results were compared to measured results. LiDAR measurements showed the closest agreement with the measured results for plant height, canopy volume, plant spacing, and row distance, outperforming both RGB and depth imaging. Under field conditions, plant spacing and row distance were estimated with accuracies of 97.5% and 94.7%, respectively, exhibiting higher measurement accuracies than RGB and depth imagery data results. Despite some discrepancies due to complex plant geometry and dynamic data collection, the results support data collection strategies critical for precision horticulture.
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
Huang Y, Huang X, Huang L, Fu S, Wu Y, Zhou D, Peng J, Liang Y, Kuang M.
ArabidopsisCottonRaman / spectroscopySeed / grain
Key messages High-accuracy NIRS models and GWAS identified a novel QTL qPO-A07-1. GhMYB86 was validated to enhance seed protein content and fiber strength, and a functional KASP marker was developed. Cottonseed is rich in protein and oil; improving its nutritional quality is vital for global food security. In this study, near-infrared spectroscopy (NIRS) models were developed for predicting cottonseed protein and oil content using least absolute shrinkage and selection operator (LASSO) regression, achieving validation R 2 of 0.969 (P -4 ). A novel stable quantitative trait locus (QTL), qPO-A07-1, was detected, within which GhMYB86 was prioritized as a candidate gene. This gene exhibited higher expression in high-protein-content varieties during ovule development. Heterologous overexpression in Arabidopsis thaliana increased seed protein content by 2.61-3.34%, whereas expression in Saccharomyces cerevisiae increased protein content by 25.81% and reduced triglyceride content by 30.72% in comparison with the control. These results demonstrate that GhMYB86 positively regulates protein content while negatively affecting oil content. A kompetitive allele-specific PCR (KASP) marker targeting a promoter A/T polymorphism revealed that the AA allele was associated with higher-protein content, lower-oil content, and increased fiber strength across both mapping and validation populations. Furthermore, the protein content- and fiber strength-favorable allele has undergone positive selection during breeding. This study provides phenotyping tools, reliable genetic resources and a molecular marker for cottonseed nutritional quality breeding, laying a foundation for the improvement in cottonseed protein content and fiber strength.
Background The quantification of fungal spores constitutes a fundamental metric in phytopathology, serving as the primary variable for inoculum standardization and being used as a proxy for disease severity. Historically, spore quantification has relied on manual hemocytometry, which remains the most precise counting process to date, where chambers such as the Malassez slide are used to count a subsample of the inoculum. However, this method applied manually is highly labor-intensive, time-consuming, and can be prone to operator-dependent variability. To overcome these limitations, we introduce MIRA (Microscopy Image Recognition & Analysis), a novel open-source software integrating You Only Look Once (YOLO) deep learning algorithms. Featuring a user-friendly graphical interface, MIRA is adaptable to multiple camera systems and supports advanced object detection models, including YOLOv11 and YOLOv26. Results We demonstrate that MIRA can be used to accurately detect and count spores from several phytopathogenic fungi, automatically measure spore surface area, and to differentiate spores across different genera. In an exhaustive comparative analysis using Pyricularia oryzae spores as an example, MIRA was benchmarked against manual gold-standard counting slides (Malassez and Kova) and indirect spectrophotometric methods (SPARK). The P. oryzae model loaded via MIRA achieved a strong correlation (R = 0.96) with manual gold standards while reducing processing time by over 90% for high-concentration samples (10⁶ spores/mL). Beyond this benchmark, we also successfully tested specific YOLO models designed to recognize macro- and microconidia of Fusarium oxysporum f. sp. cubense , a model for Pseudocercospora fijiensis , and a single multiclass model capable of identifying six different rice pathogenic fungi. We provide comprehensive tutorials for operating the software and training custom detection models for free using Roboflow and Google Colab. MIRA is available both as open-source Python code and as standalone executables for Windows and Linux. Conclusions MIRA provides a rapid, accurate, and highly reproducible alternative to manual spore counting, effectively removing a major bottleneck in phytopathology workflows. By combining advanced YOLO-based deep learning with an accessible interface and comprehensive training resources, MIRA makes accessible automated image analysis for researchers without programming expertise. Moreover, MIRA drastically improves the efficiency of high-throughput disease phenotyping and can be adapted for a wide range of microscopic quantification tasks across various biological disciplines.
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.
Accurate assessment of crop water status is critical for precision irrigation and sustainable water management in agriculture. This study develops a UAV-based thermal infrared inversion framework for high-resolution canopy temperature retrieval and irrigation decision support in tea plantations. The proposed approach integrates multi-frame image mosaicking, threshold-based canopy extraction, and a gray-temperature calibration model to generate spatially continuous canopy temperature maps. Crop water stress was quantified using the Crop Water Stress Index (CWSI), and its reliability was further evaluated by analyzing its relationship with stomatal conductance. The framework further estimates soil moisture status and irrigation requirements based on a threshold-based irrigation strategy. The results show that the linear gray-temperature calibration model achieved a maximum absolute error of less than 0.3 °C and that the calculated CWSI and estimated irrigation requirement were strongly correlated with measured stomatal conductance, with R 2 up to 0.91. The proposed method provides a practical technical workflow from UAV thermal imagery acquisition to canopy temperature retrieval and quantitative irrigation decision-making, demonstrating its potential for precision irrigation management in tea plantations.
This study investigates the automatic classification of crop phenological stages from low-altitude UAV RGB imagery. The dataset included 11,489 images of five crops: sunflower, rapeseed, soybean, wheat, and barley. The images were annotated using the Biologische Bundesanstalt, Bundessortenamt und Chemische Industrie (BBCH) scale, with labels corresponding to either single stages or stage ranges to reflect heterogeneous field conditions and transitional crop states. A pretrained ResNet18 model was adapted to the task using transfer learning. Training was conducted in two stages: first, the classification head was optimized while the backbone remained frozen; second, the entire network was fine-tuned. The model achieved strong internal test accuracy across all crops, with 100% test accuracy for rapeseed and barley, more than 99% for the remaining crops, and a mean accuracy of 99.73% under the studied survey conditions. The results also compare favorably with previously reported studies on UAV-based phenological classification. Overall, the findings support the potential of low-altitude UAV imagery and deep learning for localized phenological assessment of selected field zones in precision agriculture, while broader deployment requires validation across independent fields, seasons, regions, and survey conditions.
Identifying plant health conditions is an emerging precision-agriculture and food-security challenge, intensified by deploying deep-learning models on memory- and power-constrained edge devices. We present a TinyMLOps pipeline spanning model design, optimization, quantization, and deployment across diverse edge devices, evaluated under controlled laboratory conditions. Using a dataset derived from the PlantVillage benchmark, 39 fine-grained classes are aggregated into three superclasses: healthy leaf, unhealthy leaf, and no leaves. The resulting system therefore performs plant health-status classification and background filtering rather than diagnosing specific diseases. We train a MobileNet-based convolutional neural network jointly optimized for classification accuracy and computational efficiency, adopting state-of-the-art hyperparameter optimization (HPO) tools. Five models are selected, four from the Pareto Front and one as the biggest evaluated model during HPO, converted to LiteRT and ONNX, and evaluated at float32 and post-training int8 precision on a Raspberry Pi Zero 2 W and an STM32H743ZI microcontroller. At float32, LiteRT is 1.87–2.65× faster than ONNX Runtime on the Raspberry Pi across all five models. Relative to their float32 LiteRT counterparts, the int8 LiteRT models are 2.83–3.67× smaller on disk and 21.3–31.2% faster on the same board, at a cost in F1-score of between 0.0010 and 0.0068. On the microcontroller, only the two smallest models deploy at both precisions; for these, the fully quantized int8 variants are 4.3× faster and 3.85× smaller in MCU flash footprint than the float32 counterparts. The mid-range model fits the 2 MB flash and 1 MB RAM budget only when quantized, while the two largest models exceed it in every configuration tested.
Reproduction assets foundThe paper's plant-phenotyping measurements are based on a derived PlantVillage dataset (39 classes aggregated into three superclasses) that the authors explicitly state is openly available in their own GitHub repository, also catalogued in the AgrifoodTEF Data Space. No author analysis code or trained model checkpointsDataset · publicThe data used in this study are derived from the openly available GitHub repository available at https://github.com/FBK-OpenIoT/PlantVillage-AugNoLeaves , accessed on 1 July 2026.Open asset ↗FBK-OpenIoT/PlantVillage-AugNoLeaveslines:1355-1403Dataset · publicPlantVillage-AugNoLeaves—AgrifoodTEF Data Space Catalogue. 2025. [(accessed on 1 July 2026)]. Available online: https://dataspace.agrifoodtef.eu/asset/did:op:3091fdcf83a05784416e585e4e45ea24afaf8c925a891921a3949acd98426f65Open asset ↗did:op:3091fdcf83a05784416e585e4e45ea24afaf8c925a891921a3949acd98426f65lines:1424-1474Plant phenotyping relevance matchOpenAlex · checked 15 Sept 2026
Published7 Aug 2026ISPRS Journal of Photogrammetry and Remote SensingCited by 0 · OpenAlex ↗
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.
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.
Phosphorus (P) deficiency severely limits soybean ( Glycine max L.) productivity. This study proposed a three-stage screening framework to identify reliable traits and P-efficient genotypes. In Experiment I, percent tolerance to phosphorus deficiency (PTPD) was calculated for ten growth parameters across 98 genotypes under P-deficient and control conditions. Principal component analysis and comprehensive evaluation identified six key indicators in Experiment I, which were subsequently refined to five indicators through further analysis: SPAD at V3 and R1, photosynthetic rate at R1, shoot dry weight at R8, and seed number per plant at R8. Experiment II re-evaluated these traits using 12 contrasting genotypes under three P levels, identifying CN 15 as the most P-efficient and SN 22 as the most P-inefficient. Experiment III further revealed that CN 15 maintained superior PSII performance and exhibited a 26.2% increase in grain P-utilization efficiency under 0 µM KH 2 PO 4 treatment. This integrated framework offers a preliminary reference for screening P-efficient soybean genotypes under controlled conditions, pending field evaluation.
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.
Jiří Chuchlík · Jan Stejskal · Jaroslav Čepl · Jiří Korecký · Salvador A. Gezan · Yousry A. El‐Kassaby
Field / plotRGB / grayscaleLeafSegmentationGrowth / development / phenology
Bud flushing timing is a crucial attribute for forest geneticists and physiologists studying adaptive variation in tree species. In clonal populations such as Norway spruce seed orchards, accurate assessment of bud flushing is essential, as this attribute strongly influences survival and fitness. However, traditional bud-flushing assessment relies on visual scoring scales that are inherently subjective and susceptible to observer bias, leading to inconsistencies when multiple technicians collect data. To address these limitations, we introduce a novel quantitative phenotype derived from segmented RGB images for quantifying bud flushing that objectively measures the ratio of light-green current-year needles to the total visible needle area of each grafted ramet. By extracting and summarizing digital image data for each ramet, this approach provides a standardized and reproducible assessment of bud flushing phenology. We evaluated clonally replicated ramets in two sites across two seasons and compared this image-derived flushing ratio with conventional visual scoring and color indices. Using mixed models, we quantified genetic control and cross-orchard stability of clonal performance. The image-derived flushing ratio tracked visual-score progression but showed higher broad-sense heritability (0.54–0.70) than visual scoring (0.45–0.65) and canopy color indices (0.20–0.51). Although site means differed strongly, clonal ranking remained stable between orchards for the image-derived flushing ratio and for visual scoring, whereas color indices showed lower stability and greater uncertainty. Overall, these results support the image-derived flushing ratio as a structure-based phenotype that improves the precision and cross-site transferability of genetic evaluation of bud flushing in Norway spruce seed orchards.
ArabidopsisMicroscopyCell / cellular structureRootMorphology / geometry measurementVisualization / data management
Super-resolution microscopy has transformed our ability to visualize subcellular structures, but its application in plant biology remains challenging due to the optical complexity of plant tissues. Here, we present a detailed protocol for tau-STED microscopy (Leica Microsystems), which combines stimulated emission depletion (STED) with fluorescence lifetime imaging (FLIM) to achieve nanoscale resolution while minimizing phototoxicity. This method leverages time-correlated single-photon counting (TCSPC) to separate fluorescence signals based on their lifetimes, enhancing signal specificity and enabling the visualization of elusive subcellular compartments in Arabidopsis thaliana root tips. The protocol covers sample preparation, fluorophore selection, microscope configuration, image acquisition, and data analysis, providing a step-by-step guide to optimize tau-STED imaging for plant cell biology. By addressing the unique challenges of plant tissue imaging, such as autofluorescence, refractive index mismatches, and light scattering, this approach facilitates super-resolution imaging of intracellular structures, including the plant endoplasmic reticulum-Golgi intermediate compartment (ERGIC). This protocol is designed to be accessible to researchers with basic microscopy experience and offers a robust framework for exploring subcellular dynamics in plants with unprecedented detail. Key features • tau-STED integrates STED signals with fluorescence lifetime via phasor analysis at confocal speeds, enabling low-noise super-resolution imaging. • Morphometry analysis workflow at super resolution.
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.
Ammonium (NH 4 + ) is the primary inorganic nitrogen source for rice ( Oryza sativa L.). Substantial progress has been made in characterizing the functions of ammonium transporters (AMTs) in roots; however, the regulatory dynamics governing subcellular ammonium compartmentation after its entry into cells, particularly its vacuolar sequestration and efflux back to the external environment, remain poorly understood. This knowledge gap stems mainly from two factors: the difficulty of applying conventional detection methods at the organellar scale and interference caused by nonspecific ion adsorption to the cell wall of intact roots. To address these challenges, we present a detailed and reproducible protocol for real-time measurement of net NH 4 + fluxes in rice roots, root protoplasts, and isolated vacuoles using non-invasive micro-test technology (NMT). The protocol covers the preparation of protoplasts and vacuoles from rice roots, the configuration and calibration of the NMT system, and the step-by-step measurement of net NH 4 + fluxes at three distinct biological levels (intact roots, protoplasts, and vacuoles). By employing a unified sample preparation and measurement strategy, this protocol enables quantification of net uptake fluxes across the plasma membrane, characterization of net efflux dynamics under specific conditions, and indirect estimation of vacuolar sequestration capacity using the isolated vacuole system. Overall, this protocol provides a flexible and robust framework for studying NH 4 + homeostasis in plants and is readily adaptable to different crop species, treatment conditions, and experimental objectives. Owing to its modular design and compatibility with standard NMT equipment, it can be readily adopted by laboratories seeking to investigate nitrogen transport mechanisms in plants. Key features • Allows for testing of NH 4 + fluxes in roots, protoplasts, and vacuoles. • Applicable to plants grown under different culture systems, including Arabidopsis thaliana grown in dishes and rice grown in hydroponic systems. • Supports both long-term and transient stress treatments. • Real-time monitoring.
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 matchEurope PMC · checked 15 Sept 2026
The precise identification of microspore and pollen at the optimal developmental stages to be induced towards embryogenesis (vacuolated microspores and young pollen) is essential for induction of in vitro androgenesis in plants. Such identification is not always easy, and it is especially difficult in recalcitrant species such as Vicia faba . The present study evaluates the relationship between floral bud and anther morphometric parameters, and microspore/pollen developmental stages in V. faba using various methods including fluorescent staining, differential interference contrast microscopy, and morphometry. We measured flower bud and anther length and width, grouping them at different intervals, and performed a detailed microscopical and anatomical analysis of buds, anthers and microspores/pollen at different stages. Our results demonstrated that flower buds in V. faba exhibit complex and irregular morphologies, with considerable variation in both sepal length and shape. Furthermore, the determination of microspore and pollen developmental stages in this species is constrained by pronounced developmental asynchrony and strong genotype dependence. Although anther length measurements correlate closely with microspore and pollen developmental stages, their practical use can be challenging. Therefore, measuring flower bud length, while excluding sepals, remains the most practical criterion for routine applications. Combining this refined morphometric approach with microscopic validation appears to be the most effective strategy for improving the identification of flower buds containing microspores or pollen at developmental stages suitable for androgenesis induction in this recalcitrant legume species.
In vivo analysis of plant stress responses remains a major challenge in precision agriculture, limiting dynamic optimization of crop growth under variable environmental conditions. Fluorescence imaging enables nondestructive tracking of stress biomarkers, but its accuracy is compromised by the low abundance of endogenous signaling molecules, tissue autofluorescence interference, and limited signal penetration depth. Here, we develop a ratiometric near-infrared IIb fluorescent probe for sensitive monitoring of endogenous hydrogen sulfide (H 2 S). This nanoprobe integrates lanthanide nanoparticles with H 2 S-responsive molecular units, enabling enhanced tissue penetration and high-resolution imaging through an absorption competition-induced emission mechanism. Under abiotic stress conditions, the probe visualizes stress-induced fluctuations of H 2 S in living plants. This work establishes an H 2 S-centered strategy for plant stress visualization and provides a foundation for developing early diagnosis platforms.
Ground-level ozone (O 3 ) adversely affects rice physiology and is associated with yield reductions. This study developed a high-resolution assessment framework integrating multi-source satellite remote sensing with econometric methods to quantify the impacts of O 3 on rice production in China's primary rice-growing region-the Middle and Lower Reaches of the Yangtze River (MLYR)-from 2019 to 2023. We fused Sentinel-5P TROPOMI total ozone column (TOC) data, a harmonized multi-satellite solar-induced chlorophyll fluorescence (SIF) product (LHSIF), high-precision rice distribution maps, and ERA5 meteorological reanalysis data. In addition to SIF, we examined multiple vegetation indicators (chlorophyll content, leaf area index, and vegetation indices) to capture broad physiological responses. A bidirectional fixed-effects panel model was employed to control for spatiotemporal confounders, revealing a significant inhibitory effect of O 3 on photosynthesis (β = -1.334 × 10 -5 , p 3 concentrations would increase regional SIF by 36.36%, while a commensurate 10% reduction in annual exposure could elevate rice yields by approximately 8.4%. This spaceborne remote sensing approach provides a robust and transferable methodology for the precise regional monitoring of ozone stress and for informing targeted mitigation strategies to safeguard crop productivity.