Forest above-ground biomass (AGB) is a crucial measure of forest carbon storage and plays a critical role in global carbon monitoring. While terrestrial laser scanning (TLS) serves as a powerful tool for AGB estimation, existing methods face key limitations. These include reliance on computationally intensive individual-tree segmentation, which struggles in dense forests with overlapping canopies, or on oversimplified structural attributes, such as diameter at breast height (DBH) and tree height, or being site-specific. Here, we present a cross-site scalable and computationally efficient modeling workflow for community-level AGB estimation using multiplatform TLS-derived metrics, without requiring tree-level point cloud segmentation. Besides community-level DBH extracted via a point-cloud-segmentation-free workflow, we selected community-level canopy structure metrics, such as leaf area densities, canopy entropy, etc. to train an AGB estimation model via Random Forest. We used data from five ecologically diverse sites across East Asia, spanning a 20° latitudinal gradient, to train and cross-validate our model. The model achieved high predictive accuracy () and showed consistent performance across all sites. Compared to site-specific models, the cross-site model performed as accurate, with canopy structural features accounting for over 40% of the explained variance and improving performance by 13.89% relative to the site-specific DBH-and-height-only model. Furthermore, our evidence suggested that such improvement was indeed driven by the inclusion of canopy structural information. However, leave-one-site-out validation showed limited transferability to environmentally distinct sites, highlighting the need for broader training datasets. Overall, our workflow provides an efficient alternative for community-level AGB estimation and underscore the importance of multidimensional canopy attributes for improving forest carbon assessment across heterogeneous ecosystems.
Dissecting the temporal genetic architecture of crop growth requires phenotyping platforms that combine high throughput with fine temporal resolution. Here, we combined multi-season RGB-UAV phenotyping with temporal GWAS and QTL mapping to resolve stage-specific genetic effects on canopy height in bread wheat. An elite panel of 174 cultivars and a recombinant inbred line population of 162 lines were monitored across three growing seasons (2021-2024) in southern Italy. Digital canopy height was extracted from plot-level 3D point clouds at multiple percentile levels and validated against ground measurements (R2 = 0.89-0.98; RMSE = 0.07-0.20 m). Temporal modelling of fitted height trajectories generated both whole-height and growth-rate traits; growth-rate associations consistently preceded whole-height associations by 100-200 growing degree days (GDD), capturing the onset of genetic effects on elongation earlier than cumulative height. An apparent tillering plateau separated two phases of genetic control: an early phase linked to vernalization (Vrn) and photoperiod (Ppd) loci and a later phase linked to gibberellin-related dwarfing genes (Rht), a sequence consistent with published expression profiles. Recurrent loci converged on five chromosomal regions, and candidate-gene prioritization highlighted a small set of regulators, including DELLA and TB1-like genes. Overall, temporal UAV phenotyping, and especially the contrast between whole-height and growth-rate traits, resolves when successive developmental loci become detectable on the canopy, providing a framework to dissect the dynamic genetic control of plant height.
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.
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 mIoUD 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
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.
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.
Arabidopsis thaliana 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.
Leaf nitrogen content (LNC) is critical for crop nutrition and precision fertilization. UAV-based hyperspectral sensing enables rapid field-scale nitrogen diagnosis in winter wheat, but stage-dependent drift in the LNC-spectral relationship limits multi-temporal monitoring. To address this limitation, this study proposed a Growth-stage Deep learning model for Leaf Nitrogen Content estimation in winter wheat (GD-LNC). The core innovation of GD-LNC lies in the synergistic integration of explicit growth-stage embedding with a dynamic feature fusion gating mechanism. The growth stage was incorporated as a key parameter in the deep learning-based hyperspectral model for LNC estimation. This strategy improved the generalization ability of multi-temporal LNC estimation during the grain-filling period. This study utilized the datasets from two growing seasons of multi-factor winter wheat experiments. The datasets included UAV-based hyperspectral images and synchronous ground samples. LNC-sensitive bands were selected via stepwise projection algorithm (SPA) and shuffled frog leaping algorithm (SFLA).Subsequently, a multi-temporal dynamic feature extraction and growth-stage embedding branch was constructed. An adaptive fusion mechanism was then employed to dynamically adjust the contribution weights of each feature. These components together enabled robust estimation and mapping of LNC. Results show hyperspectral features strongly responded to LNC, and SFLA-selected bands retained key spectral information, improving estimation accuracy and stability. GD-LNC outperformed PLSR, LSTM and RF, with SFLA-GD-LNC achieving R2 of 0.88 and RMSE of 0.20% (training), and R2 of 0.85 with RMSE 0.22% (validation), and R2 of 0.88 with RMSE 0.21% (test). The model captured spatial and temporal LNC dynamics, providing a generalizable UAV hyperspectral framework for nitrogen monitoring and cross-stage adaptation.
Accurate color representation is critical for UAV-based crop phenotyping, yet UAV images are often distorted by variable illumination and camera exposure settings. Here, we propose Lite U-net FiLM (LUF-net), a lightweight U-net framework integrated with feature-wise linear modulation (FiLM) layers, which incorporates both multispectral-derived irradiance and camera exposure parameters as auxiliary inputs. This metadata-aware design enables dynamic modulation of intermediate image features, allowing the network to disentangle unified canopy color from environmental artifacts. Experiments conducted across diverse crop types (soybean, rice, maize), flight altitudes (6m, 12m, 25m, 40m), and illumination conditions (overcast skies, cloudy, sunny, early morning) demonstrate that the LUF-net model substantially reduces the mean absolute percentage error to below 5.5%, outperforming conventional Gray-world, U-net, AlexNet-MLP, and SIDBlock-MLP methods. Ablation experiments further show that both irradiance and exposure metadata provide complementary information, while FiLM-based conditioning effectively integrates these acquisition parameters into feature learning, jointly contributing to improved color reconstruction performance. Moreover, correlation analysis indicates that the model's color reconstruction errors are weakly dependent on irradiance and exposure settings, confirming that LUF-net reduces sensitivity to external imaging conditions while maintaining physically meaningful and robust corrections.
Individual tree segmentation from LiDAR point clouds is critical for forest inventory and ecological monitoring. However, accurate delineation remains challenging in complex forest environments with dense crown overlap, occlusion, and multilayer vertical structures. To address these challenges, we propose ForestSeg3D, a semantically guided framework that improves individual tree segmentation through hierarchical semantic supervision and bidirectional cross-task distillation. Specifically, hierarchical semantic supervision (HSS) introduces a coarse-to-fine semantic learning scheme with explicit cross-level consistency, where coarse supervision distinguishes Tree from Non-Tree and fine-grained supervision further decomposes scenes into Ground, Wood, and Leaf, providing structured semantic priors for instance learning and enabling more discriminative tree representations. Building on this representation, bidirectional cross-task distillation (BCTD) explicitly couples semantic prediction and instance partition, reducing conflicts between semantic labels and instance boundaries and improving delineation in crowded forest scenes. To support large-scale inference, ForestSeg3D further incorporates semantic-aware region merging (SRM), a scene-level consolidation strategy that alleviates cross-region conflicts caused by overlapping crowns and boundary-crossing trees. Extensive experiments on the FOR-instanceV2 dataset show that ForestSeg3D achieves the best overall performance in detection completeness, false-positive suppression, and F1-score, while maintaining competitive mAP. The method also maintains strong performance on the ForestSemantic dataset under three-fold leave-one-plot-out cross-validation. Improved segmentation further enables more reliable estimation of tree-level structural attributes, including tree height, crown diameter, and crown volume. These results demonstrate the effectiveness and practical value of ForestSeg3D for individual tree segmentation and downstream forest structural assessment from LiDAR point clouds.
Computed tomography enables non-destructive phenotyping of fruit internal structure but traditionally requires hundreds of projections, limiting throughput. Under sparse-view conditions, conventional and learning-based methods both suffer from streaking artifacts and regional distortions that degrade trait quantification. This study present CitrusGS, an integrated framework that achieves high-fidelity 3D reconstruction and precise morphological phenotyping of citrus fruit from only 15 projections using radiative 3D Gaussian splatting. Our method employs sparse-point initialization, optimized loss composite, and dual-stage pruning to suppress artifacts while preserving anatomically critical details with significantly higer convergence efficiency. In the citrus fruit datasets, CitrusGS achieves 29.78 dB PSNR and 0.870 SSIM, outperforming corresponding baseline method by 1.58 dB and 0.067 in SSIM, and enables automated extraction of ten external and internal phenotypic traits with R2 larger than 0.944. Moreover, the framework shows initial zero-shot transferability across pathological citrus samples and additional horticultural specimens without retraining. By reconciling acquisition efficiency with anatomical fidelity using low-cost X-ray hardware, CitrusGS provides a promising framework for high-throughput, non-destructive phenotyping in breeding and grading applications.
Accurate counting of rubber trees is important for yield estimation, refined field management, and the sustainable development of the rubber industry. In deep-learning-based counting, reliable canopy delineation is a prerequisite for accurate segmentation and subsequent statistical analysis. However, maintaining stable canopy segmentation under long-term and complex field conditions remains difficult, mainly because of overlapping crowns, interference from background vegetation, and pronounced seasonal changes. To address these issues, this study proposes CSAF, a rubber tree canopy segmentation model designed for high-precision canopy segmentation and counting from UAV imagery. CSAF consists of three main components: the Boundary Continuity Modelling Module (BCMM), the Physical Morphology Constraint Module (PMCM), and the Cross-Temporal Learning Adaptively Module (CTLAM). BCMM improves the representation of complex canopy boundaries by using the Fourier transform to suppress feature noise and state-space modelling to refine boundary responses. PMCM introduces a Poisson diffusion prior to constrain the segmentation results according to canopy morphology, thereby reducing the influence of irregular background vegetation. CTLAM adjusts the replay ratio and scale through a state-feedback mechanism, which helps mitigate inter-seasonal variation and improve cross-temporal generalization. In addition, this study presents RT-Set, a UAV-based rubber tree canopy dataset containing 5281 high-resolution images that cover the full growth cycle of rubber trees, from the budding stage to the defoliation stage. RT-Set provides a dedicated benchmark for precise rubber tree canopy segmentation and counting. On RT-Set, CSAF achieves an AP50 of 76.63%, outperforming the compared methods. Experiments on three additional public datasets further show that CSAF remains competitive with mainstream approaches, including DetecTree2 and Cascade Mask R-CNN. In the counting experiments, CSAF also shows high accuracy and stable performance, suggesting its practical value for rubber tree counting under complex field conditions. The RT-Set dataset is available at https://github.com/zengjiangquan1/RT-Set.
Three dimensional (3D) instance segmentation is essential for precision characterization of tree architecture at the branch level, which supports both tree fruit crop breeding and the development of robotic systems for orchard management. Existing methods usually use sparse convolution-based operation, which requires a coordinate quantization preprocess to generate sparse tensors, risking the loss of geometric details for fine-grained downstream phenotyping tasks. To overcome this challenge, we developed the dynamic sparse point-voxel transformer (DSPVFormer) model for the efficient and accurate 3D instance segmentation of high-resolution point clouds for dormant apple trees. Our hybrid DSPVFormer architecture maximizes the use of raw point features by dynamically mapping and aggregating the raw point features into the sparse voxel embeddings, capturing strong geometric features that may be discarded during quantization. Evaluations demonstrate that DSPVFormer achieved statistically significant improvements over baseline models on most instance segmentation metrics, which are further translated into more accurate phenotyping evaluation including branch counting and pruning map generation. These advances directly benefit downstream applications in plant phenotyping and robotic pruning for tree crops such as apples. Meanwhile, experimental results on phenotyping tasks suggested that phenotyping-specific evaluation metrics should be prioritized over upstream computer vision performance metrics to realize the full potential of high-throughput phenotyping for real-world applications.
Advancing plant phenomics requires linking high-resolution phenotypic data to plant performance under environmental stresses like grazing. However, how intrinsic biological factors, specifically individual developmental stage, mediate phenotypic trade-offs in response to management remains poorly quantified. Using a phenomics approach on the dominant grass Stipa bungeana within a two-decade grazing experiment, we integrated multi-year, individual-level trait data to dissect the effects of grazing season (cold/warm), intensity (light/moderate/heavy), climate, and developmental stage (proxied by basal diameter) on vegetative and reproductive tiller phenomes. We show that warm-season grazing increased tiller production, whereas cold-season grazing simplified phenotypic correlation networks. Notably, developmental stage was the dominant driver of vegetative growth and individual biomass, while reproductive investment was primarily governed by external drivers (climate and grazing intensity). Path modeling revealed that developmental stage indirectly enhances sexual reproduction by fueling vegetative investment, jointly determining final biomass. This study suggests that individual developmental stage as a potential internal integrator of grazing signals, reshaping phenotypic architecture. Our findings provide a phenotype-driven framework for precision grassland management. By advocating for the monitoring of developmental stage composition, we bridge phenomics with sustainable practices, enabling dynamic grazing strategies that optimize the balance between productivity and ecosystem resilience.
Sweet potato is a nutritionally valuable crop that contributes to food security, owing to its storage roots rich in starch, sugars, and antioxidants, while requiring minimal cultivation inputs. Estimating its yield based on visible above-ground traits remains challenging due to weak and inconsistent correlations between shoot biomass and storage root development. Therefore, direct assessment of underground biomass is essential. In this study, we demonstrate the field application of ground penetrating radar (GPR) for non-destructive detection and yield estimation of sweet potato. GPR is a geophysical technique that typically transmits ultra high frequency radio waves into the soil and records reflections from subsurface objects. Electromagnetic wave simulations within the soil-root system revealed GPR signals that strongly correlate with root length, forming the basis for yield quantification. We developed an image-processing pipeline comprising static correction, gain adjustment, noise filtering, and hyperbola segmentation via the Hough transform to enable semi-automated storage root detection from GPR data. By integrating detection and quantification approaches, a linear regression model predicting sweet potato yield from GPR signals achieved moderate accuracy (R 2 = 0.567, normalized RMSE 0.190). We established a non-destructive and low-labor approach for monitoring root systems, providing a foundation for rapid, scalable, and field-ready yield estimation in sweet potato and other root and tuber crops.
Soybean (Glycine max (L.) Merrill) is a highly important crop widely used for food, edible oil, animal feed, and microbial fermentation products. Traditional phenotypic measurement methods are often time-consuming, labor-intensive, destructive to plants, and prone to human error. High-Throughput Phenotyping (HTP) enables precise assessment of multiple soybean phenotypic features, including morphology, physiology, diseases, pests, and agronomic traits. Artificial Intelligence (AI) is a research field dedicated to developing algorithms for multiple tasks. This review highlights the application of HTP and AI in soybean breeding programs. We discuss the challenges of implementing HTP in soybean breeding and focus on the potential and limitations of Deep Learning (DL) to support soybean breeding goals. We demonstrate the application of HTP to key soybean traits, several HTP platforms, as well as DL applications across different datasets and strategies for developing large foundation models. While integrating AI into soybean breeding programs remains a challenge, leveraging HTP data and Large Language Models (LLMs) could reshape soybean breeding.
Leaf inclination angle (LIA) strongly affects canopy photosynthesis and is an important trait for crop productivity. Difficulty in determining mean LIA throughout the canopy makes it a trait rarely evaluated in crop science research. Here, we report a novel, low-cost, simple method to rapidly measure the mean LIA throughout the canopy. We captured RGB and depth images of the crop canopy with a compact stereo depth camera aligned parallel to the ground in a handheld manner. The mean LIA was estimated through simple vector calculations. To validate the accuracy of this method, we used tests with a white board fixed at known angles and single leaves whose angle had been measured with a digital level. The mean LIA was strongly correlated with the mean tilt angle—a trait similar to LIA that was measured with a plant canopy analyzer, an expensive instrument for measuring leaf area index. Using our method, we were able to detect differences in LIA among cultivars, growth stages, and plant densities in rice, barley, soybean, and wheat. Because this method is very easy and has minimal financial, time and labor requirements, it is likely to be highly beneficial for crop scientists whose primary research is conducted in fields.
Rice seedling salt-tolerance evaluation commonly relies on visual scoring or destructive assays, which are subjective, labor-intensive, and difficult to standardize for population-level analysis. This study developed a new deep-learning-assisted hydroponic RGB phenotyping framework for standardized salt-stress evaluation and genetic mapping in rice seedlings. The framework integrates controlled hydroponic cultivation, RGB imaging, RicePhenoSeg-assisted annotation and trait extraction, ELMERF-based semantic segmentation, and image-derived quantification of salt-induced shoot injury. Using this framework, we constructed the Rice Seedling-Salt RGB Dataset (RSSD), which contains green shoot tissues, yellow shoot tissues, roots, and background from hydroponically grown rice seedlings. Based on RSSD, ELMERF achieved a mean Intersection over Union of 51.4% and a mean Accuracy of 89.5%, outperforming nine representative segmentation models. We further defined shoot yellowing rate (SYR) as an image-derived quantitative trait describing visible salt-induced shoot injury. The framework was applied to 261 re-sequenced rice accessions for population-level phenotyping and genome-wide association analysis. Compared with standard evaluation score and seedling death rate, SYR showed a more continuous phenotypic distribution and detected 36 significant SNPs, including a major signal near the Saltol/OsHKT1; 5 region. Notably, 34 SYR-associated SNPs were not detected by conventional visual scores. Overall, this study provides a targeted hydroponic RGB phenotyping framework for standardized rice seedling salt-stress evaluation and genetic analysis.
Accurate pest detection is crucial for safeguarding crop productivity and quality. Few-shot image classification offers a promising solution for automated pest recognition in limited-data settings. Recent studies have demonstrated the effectiveness of meta-learning with episodic tasks, enabling rapid adaptation to new classes. However, due to the limited number of samples available for each task, pretraining an appropriate embedding network for meta-learning plays a critical role in determining the few-shot classification performance. To enhance feature extraction, recent approaches have adopted supervised whole-classification training as a preparatory step before meta-learning. Nevertheless, when the labeled data for whole-classification training is insufficient, the performance improvements become limited. To address these challenges, this study proposes a method called MoCoProto, which combines self-supervised learning with meta-learning. The proposed method first utilizes momentum contrast-based self-supervised visual representation learning to pretrain an embedding network using unlabeled images. The pretrained embedding network is then applied to a prototypical network for few-shot image classification through meta-learning. For self-supervised pretraining, a broad-domain agricultural dataset comprising 357,705 images from 65 pest species and 144 plant diseases was used, from which all label information was excluded, relying solely on the image data. Subsequently, meta-training and final evaluation were conducted using the IP102 insect pest dataset, and the experimental results demonstrated that the proposed MoCoProto achieved the best performance, recording accuracies of 49.33% in 1-shot and 70.13% in 5-shot classification.
Accurate assessment of plant dry matter (PDM) and plant N accumulation (PNA) provides essential indicators for precision nitrogen (N) management in rice production. However, purely data-driven models struggle to generalize due to the spatial scarcity of ground-truth physiological data. To address this, a physiology-informed long short-term memory (PI-LSTM) framework was developed for robust regional N diagnosis and variable-rate fertilization. First, the model was pretrained to internalize crop growth dynamics using a DSSAT-based simulation library, which spanned 2000 representative fields and 700 management scenarios to provide physiologically consistent pseudo-labels. Subsequently, the framework was fine-tuned using multi-year field observations (2020, 2023, 2024), Sentinel-2 time-series data, and meteorological inputs. The proposed LSTM framework outperformed conventional machine learning approaches in estimating PDM and PNA, achieving five-fold cross-validation R2 values of 0.87 and 0.83, respectively. Based on these biophysical estimations, the N nutrition index (NNI) diagnosis achieved a 67.3% overall classification accuracy. Furthermore, by integrating the critical N dilution curve, the critical PNA and accumulated N deficiency (AND) were quantified, which served as the basis for developing the AND-based N recommendation algorithm (ANDA). Finally, variable-rate topdressing field experiments conducted across seven sites in 2024 and 2025 demonstrated that the ANDA reduced N input by 13.4% compared with farmers' practices, while maintaining or increasing yield and improving N partial factor productivity by 18.6%. This study provides a reliable, physically consistent decision-support framework for regional-scale precision N management.
