Escalating climate fluctuations and the increasing frequency of compound extreme weather events pose severe threats to global food security. Current operational crop yield forecasting systems, which predominantly rely on process-based models or traditional statistical approaches, often underestimate yield losses during climatic extremes due to their inability to capture complex, non-linear climate-yield dynamics. While deep learning (DL) offers a transformative alternative, its widespread adoption remains limited by its “black-box” nature and the difficulty of processing long agro-meteorological time series without losing critical signals from climatic anomalies. To address these gaps, we propose CropFusionNet, a novel and interpretable architecture inspired by the Temporal Fusion Transformer (TFT), designed for district-level (NUTS-3) crop yield forecasting. CropFusionNet integrates daily time-varying climatic variables with static agro-environmental covariates and offers a more interpretable deep learning framework, enabling attribution of predictions to key environmental drivers, though not fully mechanistic. Applied to Germany's principal crop commodities, CropFusionNet consistently outperformed established deep learning baselines and operational frameworks such as MARS and ABSOLUT. It achieved high predictive accuracy for silage maize (R2 = 0.75; MAPE = 8.66%), winter wheat (R2 = 0.60; MAPE = 8.06%), and a moderate accuracy for winter barley (R2 = 0.45; MAPE = 10%), successfully capturing inter-annual variability and accurately reproducing extreme negative yield anomalies during severe drought years (2003, 2018, and 2022). Lead-time analyses further demonstrated that reliable forecasts for winter cereals (wheat and barley) can be made 40–60 days before harvest, whereas spring crops (silage maize) can be predicted up to 70 days in advance. By disentangling key environmental drivers and effectively representing non-linear system dynamics, CropFusionNet represents a fundamental transformation towards scalable, accurate, uncertainty-aware, and interpretable agricultural forecasting, providing essential support for proactive food security management. The model implementation is available at https: //github.com/geonextgis/CropFusionNet.
The unchecked and rapid proliferation of weeds poses a substantial threat to Paeonia lactiflora Pall. (PLP) cultivation. Weed management in PLP fields currently relies primarily on manual weeding, which is labor-intensive, costly and inefficient. Intelligent weeding is a key enabler for advancing the Chinese medicinal herb industry toward sustainable production. However, field environments exhibit numerous uncontrolled factors, such as complex backgrounds, occlusion and overlap among weeds, and lighting variability. It is difficult for single-modality RGB vision systems to achieve precise and robust weed detection. To address these challenges, a depth-guided, single-stage RGB-D weed detection Transformer, termed FW-DETR, was proposed. The framework aims to efficiently utilize spatial geometric priors to guide and fuse visible features. FW-DETR adheres to the modality-graded feature processing paradigm, and designs a Cross-Modal Hierarchical Feature Interaction module (CM-HFI) to form the RGB-D feature encoder, which implements multi-scale bimodal feature interaction and fusion in complex scenes. The proposed Cross-Modal Global Context Modeling module (CM-GCM) enhances spatial awareness and global context capabilities, whereas the Spatially Adaptive Local Feature Modeling module (SA-LFM) recalibrates RGB features and refines local details. The developed Cross-Modal Cascaded Attention module (CM-CAM) incorporates spatial information into high-level features, improving the representation of critical region features. Subsequently, FW-DETR was integrated into the vision system of weeding robot to generate robot-executable 3D spatial operation points. Experimental results show that FW-DETR exhibits good robustness and generalization across diverse field environments, achieving 84.4% mAP0.5 and 51.7% mAP0.5:0.95. The mean pixel positioning errors (Δx,Δy) of the weeding operation points are 2.57 and 2.61, respectively. Field trial results indicate that the weed removal rate and PLP damage rate are 89.5% and 2.3%, respectively. These results provide a reliable foundation and practical guidance for the development of intelligent weeding equipment for Chinese medicinal herbs.
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 R2 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.
Precise detection and segmentation of tomato pedicels are critical perception tasks for automated harvesting by picking robots. Addressing challenges posed by complex orchard environments and loss of small target features, this paper constructs a tomato pedicel dataset and proposes the YOLO11-FFTDA model. First, the backbone network incorporates a CFP module featuring frequency-domain attention (FFTAT) to explicitly enhance high-frequency feature extraction for slender stems. Second, the integrated C2PSADA module employs deformable attention to adaptively fit stem shapes, effectively suppressing background noise and occlusions. Third, a small object detection branch is added to improve capture capability for minute targets. Experimental results demonstrate that the YOLO11-FFTDA model achieves 88.1% precision, 76.6% recall, 80.0% mAP50, and 32.9% mAP50–95. Compared to the baseline YOLO11 model, its mask quality metric mAP50–95 improves by 3.0 percentage points and compared to the lightweight Yoloact model, mAP50 and mAP50–95 increased significantly by 19.9% and 9.8%, respectively. Five-fold cross-validation revealed extremely low standard deviation, validating the model's high stability and robustness against data variance in complex orchard scenarios. This method enables fast and accurate tomato stem segmentation in computationally constrained complex orchard environments, providing reliable technical support for the visual servo system of harvesting robots.
Robotic grading of fruit requires reliable contact-based firmness sensing, as firmness is a key indicator of maturity and eating quality but is often difficult to infer from vision alone. This study presents a robotic tactile sensor integrating electrical impedance tomography (EIT) for non-destructive kiwifruit firmness estimation. A thin, compliant hydrogel sensing sheet with a peripheral 16-electrode array is integrated into the fingertips of a parallel-jaw gripper to acquire EIT measurements during robotic grasping. A dual-branch spatiotemporal fusion network was developed to jointly exploit reconstructed EIT images and voltage sequences, with cross-branch attention and adaptive fusion for grasp-based firmness prediction. Experiments were conducted on 366 kiwifruits from three cultivars, with stiffness (S), skin firmness (Fs), and flesh firmness (Ff) obtained from puncture-based reference measurements. On the held-out test set from the pooled multi-cultivar dataset, the model achieved an MAE of 0.27 N mm−1, an RMSE of 0.35 N mm−1, and an R2 of 0.95 for S. Ablation experiments showed that the two sensing branches provided complementary information and that the cross-branch attention module further improved prediction performance. Leave-one-cultivar-out (LOCO) experiments further indicated promising cross-cultivar transferability within kiwifruit. These results demonstrate the feasibility of EIT-based tactile imaging for non-destructive robotic kiwifruit firmness estimation and highlight its potential for postharvest grading applications.
Grasshoppers pose a serious threat to agriculture, livestock and grassland ecosystem, necessitating accurate risk warning. Existing methods often overlook grasshopper ecological mechanisms and the impacts of class imbalance, while lacking combinatorial optimization across machine learning (ML) pipelines, which weakens interpretability, accuracy and performance. To address these challenges, this study proposes GrasshopperML, a mechanism-guided ML framework with pipeline-level combinatorial optimization, integrating multi-source remote sensing and meteorological data. Its three key innovations are: 1) a high-dimensional environmental feature set reflecting overlap and spatiotemporal heterogeneity of grasshopper developmental stages (GDSs), incorporating precipitation, temperature, vegetation, soil, and topography for ecologically grounded modeling; 2) GDSTree, a new mechanism-guided feature selection (FS) algorithm that identifies GDS-specific key factors to enhance ecological interpretation and prediction accuracy; 3) a pipiline-level combinatorial optimization strategy exploring optimal combinations across data preprocessing, FS, and prediction modeling. To tackle class imbalance, the framework evaluates 40 combinations of five adaptive under-sampling (AUS) methods and eight ML algorithms. Experiments based on 2022 field survey data from Hulunbuir, China, show the optimal AUS-GDSTree-MLA combination outperforms five advanced FS algorithms, achieving a 62% feature reduction, 90.83% balanced accuracy, and 88.40% prediction performance. A 2023 risk map further validates GrasshopperML's practical utility, with 85.19% of presence sites falling within predicted moderate/high risk areas. This study provides an AI-based solution that integrates ecological knowledge, mechanism-guided FS, and ML pipeline optimization for advanced pest risk warning.
All-day detection and occlusion scenario analysis are critical vision techniques to ensure the continuous operation of tomato-picking robots. Most detection algorithms struggle with varying weather and light. To improve all-day tomato detection accuracy and robustness, this study introduces a lightweight, high-precision method combining deep learning and image processing. At night, HSV color distances between mature tomatoes and backgrounds were calculated under different light, creating an all-day image dataset. An improved model EMSYOLOv8 based on YOLOv8n introducing EfficientViT network, MCA attention mechanism, and SIoU loss function was trained to accurately detect and locate tomatoes under various conditions. Within the detected bounding boxes of mature tomatoes, image processing techniques such as dynamic OTSU, Hough detection with filtering, and connected component analysis were employed to analyze occlusion scenarios and differentiate between occlusion types (leaf-occluded, branch-occluded, and couple-occluded). This AI vision technology for picking robots provides valuable guidance for optimizing picking strategies. Experimental results show that the improved YOLOv8 model achieved 91.9% mAP on the all-day tomato dataset, representing improvements of 1.2, 1.8, 2.5 and 2.8 percentage points compared to YOLOv11, YOLOv10, YOLOv8, and YOLOv5 models, respectively. For occlusion scenario analysis, the overall accuracy of occlusion type determination reached at least 82% under couple-occluded. Feasibility tests in greenhouse environments demonstrated that the proposed method enabled the robot achieve a single-fruit picking success rate of 80% without occlusion under natural light, and 72% under couple-occluded interference; the single-fruit picking success rate reaches 73% without occlusion under nighttime supplemental light, and 63% under couple-occluded interference, which providing robust visual technology support for the development of all-day tomato-picking robots.
In agricultural artificial intelligence image classification tasks, existing models often exhibit redundant feature extraction and limited generalization capability. To address these practical demands and technical challenges, we propose a Cognitive-Inspired TriPercept CrossFusion Self-Distillation Network (CTNet). First, to counteract the adverse effects of environmental variations and motion blur inherent in agricultural image acquisition, we introduce a comprehensive data augmentation strategy specifically designed for agricultural imagery, thereby enhancing model robustness. Second, we present an efficient and lightweight TriPercept CrossFusion Network (TPCF), which concurrently captures multi-branch representations and emphasizes discriminative local features. A cross-branch fusion mechanism facilitates inter-branch feature interaction, enabling effective feature learning with minimal computational overhead. Third, we devise a novel end-to-end trainable self-distillation framework, termed Cognitive-Inspired Multi-Stage Self-Distillation (CISD). CISD operates without external pretrained teacher models; instead, it performs intra-model knowledge transfer across layers and employs a composite objective combining classification loss (cross-entropy), distillation loss (KL divergence), and hint loss (feature map alignment). Inspired by cognitive load theory, the framework dynamically adjusts the weighting of these loss components to achieve cognitive-load-aware multi-stage self-distillation, thereby improving generalization. We evaluated CTNet on a custom lychee disease dataset and six public benchmarks. On our augmented lychee dataset, CTNet achieves 99.03% accuracy. On three normal-collected datasets (PlantVillage, CottonWeedID15, and PotatoLeafDisease) the model attains accuracies of 99.74%, 91.81%, and 99.52%, respectively. On three controlled-collected datasets (PlantSeedling, RiceImage, and SoyBeanSeed) the corresponding accuracies are 97.45%, 99.97%, and 93.08%. CTNet consistently outperforms state-of-the-art classification methods across all benchmarks, demonstrating its effectiveness as a practical solution for agricultural image classification tasks.
Weed diagnosis and management in agricultural fields still rely heavily on manual observation and expert knowledge, resulting in limited efficiency, accessibility, and decision interactivity. Although computer vision enables automated weed detection and large language models (LLMs) provide new possibilities for intelligent decision support, an integrated framework that connects visual perception with management reasoning is still lacking. To address this challenge, this study proposes an integrated framework for intelligent weed diagnosis and management that combines a weed detection model, an LLM-based decision module, and a user-oriented diagnosis system. A detection model named Mamba-Weed is developed based on the Mamba object detection architecture, incorporating a Lightweight Adaptive Extraction (LAE) module and a Semantics and Detail Infusion (SDI) module to enhance discriminative feature learning and multi-scale feature fusion. An intelligent weed management decision module is then constructed using the DeepSeek LLM, together with Retrieval-Augmented Generation (RAG) and structured prompt engineering to generate domain-specific management recommendations. Experiments on the Weed25 dataset show that Mamba-Weed achieves 92.10% precision, 89.50% recall, and 93.80% mAP@50, demonstrating strong detection performance in complex field environments. The proposed framework establishes a unified pipeline from visual weed identification to intelligent decision-making, providing a practical solution for real-time field weed monitoring and scalable precision weed management.
Off-type plants in hybrid rapeseed compromise genetic purity, reduce seed yield and uniformity, and may jeopardize certification. Maintaining high-purity seed lots therefore requires early, reliable identification and removal of off-types (i.e., roguing) at field scale. We propose a fusion-based, lightweighted deep learning approach, which is designed for on-robot deployment, to enhance the detection performance of parental rapeseed plants by integrating CloFormer and DySample (CD) modules. Based on the CNN detections, we develop a row-wise roguing strategy, along with plant tracking and counting rapeseed plants. A roguing system is designed for automatic and constant hybrid seed production usage, which consists of two main components: the sensing-control module and the roguing module. An RGB camera, edge computing unit, and PLC connected to motor drivers provide continuous imagery and real-time feedback to the removal module. The motor triggers blade cutting of rapeseed plants via a crank-rocker mechanism. An electric four-wheel steering robot served as a mobile platform to install and connect the removal system via CAN-bus. Test results show that the proposed CD model outperforms classical models in precision, recall, and F1-score, while maintaining a smaller model size. The roguing system responded effectively when integrated with the CD model and the off-type removal strategy. The identification accuracy of the parental rapeseed plants achieved a deviation of only 2 out of 32 plants. Field test results demonstrated that the plant count closely matched the real one. Using the robotic roguing system, the correct removal rate exceeded 91%, with missed and mis-pruned rates both below 2%.
Leaf spot disease, as a pervasive foliar disease, has become a significant factor limiting stable grain production. Rapid, accurate remote sensing of disease status supports precision prevention and control. Sun-induced chlorophyll fluorescence (SIF) is closely coupled to photosynthesis and highly sensitive to disease stress, but the intertwined physiological and non-physiological influences impede mechanistic attribution and robust quantification of SIF variability, and SIF alone cannot represent the full cascade of stress damage. To overcome these constraints, SIF was decoupled using three reflectance-based approaches (NIRvP, FCVI×PAR, NIRvR) and radiative transfer model simulations, enabling separation of physiological and non-physiological components; Lindeman-Merenda-Gold (LMG) relative importance decomposition further identified the dominant control on SIF variation. In parallel, three-band vegetation indices (TBIs) tailored to stress-specific spectral signatures were developed, and compact composite indices were formed by algebraic coupling of SIF and its derivatives (ΦF−p,SIFyield, SIFtotal and SIFmea) with TBIs through addition, multiplication, and exponentiation; four regression families (linear, quadratic polynomial, logarithmic, and exponential) were evaluated for DI retrieval. With escalating disease severity, the contribution of ΦF−f increased monotonically from 51.501% to 89.444%, indicating that disease-driven physiological modulation is the principal source of SIF variability. Four TBIs, namely TBI-3 (554, 890, 710), TBI-7 (762, 866, 702), TBI-11 (758, 878, 542), and TBI-16 (918, 878, 542), exhibited strong correlations with the disease index, reaching 0.8678, 0.8674, −0.8423, and − 0.8393, respectively. Composite indices combining the two parameter families further strengthened DI relationships; quadratic fits for ΦF−p∗∗TBI−3, ΦF−p∗∗TBI−7, SIFyield∗∗TBI−3 produced R2 minima of 0.800, 0.815, and 0.831 and maxima of 0.832, 0.836, and 0.841, with narrow spreads confirming robustness to data partitioning. Overall, these findings provide a mechanistic perspective on SIF variability in maize leaf spot and demonstrate the potential of integrating SIF-derived parameters with three-band vegetation indices for disease severity quantification under field conditions.
Plant height is a key 3D phenotypic trait for assessing crop growth, biomass accumulation, and lodging resistance. To overcome the practical limitations of conventional plant height measurement methods, this study proposes a novel vision foundation model-based framework named Depth for Plant Height (Depth4PH) for plant height estimation in agricultural scenes using low-cost monocular RGB imaging. As part of our contributions, a synthetic–real coupled multimodal dataset was constructed by integrating Blender virtual agricultural scenes (Blender VAS) with real field images. Building upon the existing Depth Anything V2 foundation model, we developed a novel module called Transfer-based Agricultural Metric Depth Anything V2 (TAM-Depth V2) for absolute metric depth estimation through parameter-efficient fine-tuning, depth decoder reconstruction, and joint loss optimization. Furthermore, we designed a novel multi-source prompt-based segmentation framework, MSP-SAM2, to generate positive and negative prompts for zero-shot crop instance segmentation. Finally, a new inverse physical plant height estimation algorithm, RANSAC-Per, was introduced to estimate plant height by combining truncated percentile statistics with local RANSAC micro-plane fitting, thereby reducing the effects of depth noise and field microtopographic variation. The result showed that TAM-Depth V2 achieved stable absolute depth estimation, with an RMSE of 0.1162 m and an AbsRel of 4.25%. Compared to the original box-prompted SAM 2, MSP-SAM2 achieved a 4.4% improvement in mIoU, reaching 91.6% and a recall of 93.2%. On a 350-plant multi-crop test set, Depth4PH achieved R2 = 0.948, RMSE = 12.23 cm, and MAE = 8.82 cm, and MAPE =10.15%, with crop-specific RMSEs ranging from 3.99 cm (cucumber) to 20.73 cm (maize), significantly outperforming the traditional Global-MinMax baseline (which had an RMSE of 22.62 cm). These results indicate that Depth4PH provides a promising foundational pathway for high-throughput crop phenotyping. With future optimization for edge deployment, it holds significant potential to support high-throughput monitoring in precision agriculture.
Farmland is an unstructured and dynamic environment, where real-time detection and avoidance of moving obstacles are essential for the safe and autonomous operation of agricultural machinery. The Dynamic Window Approach (DWA) has been widely used for obstacle avoidance because of its high computational efficiency. However, when fixed evaluation parameters are used, DWA has limited adaptability in complex dynamic environments, is prone to local optima in multi-obstacle interaction scenarios, and often produces long avoidance paths with insufficient trajectory smoothness. To address these limitations, this study proposes a SAC-DWA based dynamic obstacle avoidance method, in which the Soft Actor-Critic algorithm is used to adaptively optimize the weight parameters of the DWA evaluation function online according to environmental changes. The proposed method jointly considers obstacle motion state, relative distance, vehicle speed, trajectory smoothness, and avoidance path length, so as to improve obstacle avoidance efficiency and path smoothness while ensuring safety. In addition, a multi-target tracking system is constructed by fusing 3D LiDAR and RGB camera data to provide accurate target tracking and state estimation for dynamic obstacle avoidance. Simulation and field experiments show that, compared with the conventional DWA method, the proposed SAC-DWA method reduces obstacle avoidance time by more than 13% and shortens path length by over 10%. These results demonstrate that the proposed method can effectively improve dynamic obstacle avoidance performance while maintaining safety, showing good potential for practical application in agricultural machinery.
Accurate segmentation of leaf lesions under field conditions is fundamental to intelligent crop health management but remains constrained by scarce annotations and large distribution shifts. In this paper, we present AgriGenSeg-Leaf, a generative data engine that transforms synthesis from a static preprocessing step into a closed-loop, performance-guided process. By integrating physiology-aware mask-first generation, physics-informed rendering, and multi-level optimization with downstream segmentation loss, AgriGenSeg-Leaf learns to produce synthetic image-mask pairs that target the empirically observed weaknesses of model. This synergy of venation-constrained geometry and photometric realism yields biologically faithful, controllable samples, improving both boundary fidelity and out-of-domain robustness under ultra-low-label conditions. Evaluations across U-Net, DeepLabv3+, and SegFormer backbones demonstrate consistent gains in Dice, mIoU, and Boundary-F scores without increasing inference cost. Deployed on Jetson AGX Orin, the model sustains real-time inference, offering a scalable pathway toward field-ready, data-efficient plant disease perception through generative intelligence.
Accurate prediction of foxtail millet yield is essential for effective field management and high-throughput breeding. Despite advances in UAV-based yield prediction for major crops, existing studies predominantly rely on single-temporal features (SFs) extracted at noon, overlooking significant diurnal dynamic signals that characterize crop responses to water stress. To address this research gap, we propose a novel approach utilizing diurnal cross-temporal features (CFs) derived from UAV-based multispectral and thermal imagery to enhance yield prediction accuracy under different irrigation regimes. During the flowering and grain-filling stages, UAV images were acquired across eight time slots (T1–T8) within a single day to capture the complete diurnal trajectory of canopy physiological responses. SFs were extracted at each time slot, and CFs were derived through summation, averaging, and range operations across multiple slots. A systematic four-step workflow was developed to determine the optimal UAV flight frequency and timing by balancing prediction accuracy with operational costs. Three ensemble learning algorithms (Random Forest (RF), Adaptive Boosting (AdaBoost), and Extreme Gradient Boosting (XGBoost)) were evaluated using multiple feature sets incorporating SFs, CFs, and their integration. Results demonstrated that CFs more comprehensively captured dynamic crop responses to water stress than SFs. Canopy features from afternoon combinations generally exhibited stronger yield correlations than morning combinations. The [T5, T8] combination was identified as optimal, providing a practical balance between prediction accuracy and operational cost. Model comparison revealed that RF exhibited greater robustness across different water treatments, whereas AdaBoost achieved higher accuracy on the test set. Feature importance analysis confirmed the dominance of CFs, with ∑VSWI ranking first across both models and growth stages. This study provides a systematic framework for utilizing diurnal dynamic signals in crop yield prediction, offering new methodological insights for precision agriculture and high-throughput phenotyping of foxtail millet and other dryland crops.
Rice lodging seriously hinders the working efficiency of combine harvesters, and it often causes mechanical blockage and harvest loss. It is very important for the combine harvester to perceive the direction of rice lodging in real time and precisely to enable adaptive control. However, the existing visual perception methods are mainly based on a discrete classification paradigm. They suffer from large quantization errors and boundary ambiguity. ST-LodgeDirMamba, which is a physics-prior-guided lightweight network, is proposed to cast the lodging direction detection into continuous regression. The method has three merits: (i) an edge-aware inpainting strategy for building a physically consistent dataset; (ii) a Vision Mamba backbone which explicitly injects the global direction priors derived from Structure Tensor via Direction Modulation and Multi-Direction Scan mechanisms to capture the complex morphologies; and (iii) a geometry-based circular loss function to solve the periodic prediction ambiguity at the boundary 0°/360°. Experiments show that ST-LodgeDirMamba can achieve the Circular MAE and Circular RMSE as low as 2.86° and 3.66°, respectively, on the test set. The Coefficient of Determination (R2) is 0.94. Zero-shot evaluations on newly annotated independent external regional datasets from Shehong, Sichuan and Quzhou, Zhejiang demonstrate that the model maintains robust generalization under severe geographic domain shifts. On the two external datasets, the model achieved MAE values of 3.84° and 4.06°, with R2 values of 0.92 and 0.91, respectively. After the TensorRT mixed-precision optimization, the model can achieve the real-time inference speed of 18 FPS on the NVIDIA Jetson Xavier NX edge platform. The model has only 7.72 M parameters. This model provides a new solution for the rice lodging direction detection and provides robust data support for the vision-based automated harvesting system. It facilitates the development of the unmanned agricultural machinery.
The brown planthopper (Nilaparvata lugens) is one of the most destructive pests of rice and poses a threat to yield stability and food security. Although host-plant resistance is the most sustainable strategy for BPH management, conventional resistance phenotyping remains labor-intensive, destructive, and poorly suited for large-scale breeding. Here, we combined hyperspectral reflectance profiling of 50 rice varieties with an interpretable machine learning framework to enable non-destructive prediction of resistance phenotypes. Using post-infestation spectral profiles, we established classification models that captured resistance states shaped by constitutive traits and inducible defense responses. Among 13 evaluated algorithms, a radial basis function support vector machine achieved the best performance on full-spectrum data within the sampled variety panel, with an average accuracy of 0.939 ± 0.015 and a maximum of 0.972. Predictive wavelengths were concentrated in the green, red-edge, and near-infrared regions, corresponding to variation in pigment regulation, canopy structure, and water status. Spectral and network analyses showed that resistant genotypes exhibited more complex but less stable spectral co-occurrence networks, consistent with physiological trade-offs associated with defense. We also tested whether resistance could be predicted before pest infestation. Pre-infestation spectra retained significant predictive power, with accuracies of 0.572 ± 0.021 for five-class classification and 0.667 ± 0.021 for binary classification, indicating that constitutive defense-associated physiological states are optically detectable before visible damage occurs. Together, our results show that hyperspectral reflectance encodes both inducible responses after infestation and constitutive defense baselines present beforehand. This work establishes a scalable, non-invasive phenotyping strategy for early resistance screening within evaluated germplasm panels, while future validation across independent and variety-level held-out populations will be required before broader deployment.
