Agricultural crop–weed visual models are commonly evaluated within a single dataset, whereas practical model selection requires evidence across datasets, annotation policies, image conditions, and inference constraints. We present a reproducible reliability benchmark for weed-presence classification, crop/weed detection, and semantic segmentation using CropAndWeed and PhenoBench as primary datasets, with CWFID and CWD30 providing external diagnostic evaluation. The study combines an auditable, task-aligned annotation-harmonization protocol with bidirectional transfer and retention analysis, calibration and selective prediction, controlled image-condition stress, uncertainty, seed stability, subgroup analysis, and batch-1 inference reference. Evaluation across these dimensions revealed direction-dependent transfer and separated models with similar source-domain performance according to confidence quality and image-condition sensitivity. In a representative detection setting, a source-domain mAP50–95 of 0.669 corresponded to a target-domain score of 0.409, equivalent to 61.1% retention, while Gaussian blur radius 7 reduced the source-domain score to 0.298. Post-hoc temperature scaling improved ECE and Brier score in 8 of 12 classification cases. External diagnostics further identified visual, annotation, and task differences that materially affect transfer. The resulting evidence base supports agricultural model selection using transfer behavior, confidence quality, stress sensitivity, and computational cost alongside source-domain accuracy.
The integration of autonomous driving technologies into tractors marks a major step forward in modern farming, improving efficiency and precision. However, slippery terrains like paddy fields present challenges that require reliable path-following algorithms to handle uncertainties and nonlinear dynamics. In this study, an enhanced path following control methodology is proposed through the integration of feedforward compensation with model predictive control (FFC-MPC) for autonomous tractors operating in low-adhesion paddy fields. The proposed algorithm integrates soil-induced lateral slip effects to adjust the wheel steering angles, with the objective of minimizing trajectory tracking errors. To evaluate the performance of the proposed algorithm, comparative assessments were conducted within a multi-body dynamic simulation framework. Additionally, field tests were performed to evaluate the performance of the algorithm in the real-world operating conditions. The results demonstrate that, during straight-line trajectory tracking, the proposed FFC-MPC algorithm achieves a 64.6% reduction in tracking error relative to the model-independent pure pursuit control (PPC) algorithm and a 35.5% reduction relative to the model-based Linear Quadratic Regulator (LQR) approach. In more complex scenarios involving curved and transitional path segments, the FFC-MPC algorithm achieves an 86.8% reduction in tracking error compared to PPC and a 55.8% relative to LQR. The performance enhancements improve the efficiency and stability of autonomous tractors in paddy fields and show potential for broader application in other high-slip environments.
In this manuscript, we present a method to estimate the effect of a treatment or intervention when outcomes are observable for a group of treatable units, but it is unknown which units received the treatment. In particular, we consider the example of agricultural yield. We consider this method especially valuable for farming studies, where publicly available information often includes data on yields, but not on particular treatments applied to different units, such as which farms used fertilizer or other yield-enhancing technologies. The method is also relevant in settings where aggregate effects are sought for individually sensitive treatment variables. Here, we devise a method that uses a k-nearest-neighbors approach to estimate the effect of an intervention when the treated units are unknown, without attempting to estimate those labels. The method requires at least two snapshots of data at different times, with at least one snapshot having a different number of treated units relative to the others. The method estimates the magnitude of the treatment effect using the change in variance associated with the application of the treatment, and is robust to seasonality effects. If the prevalence of treatment is unknown, then the method estimates a lower bound. We present use cases with real data and test the method on synthetic data where ground truth is available. The method successfully recovered treatment effects within confidence intervals when effects were relatively homogeneous, including a fertilizer RCT and subgroup drought analyses, while performance deteriorated when treatment effects were heterogeneous.
Caged-layer farming is the dominant mode of automated layer production, making precise production monitoring essential. However, current methods rely on detached, single-task models that cause computational redundancy and lack data synergy. Furthermore, these isolated systems struggle with dynamic occlusion on conveyors and fail to provide the granular traceability required for individual cage management. To address these challenges, this study proposes an integrated monitoring framework that synergizes instance segmentation, machine learning, and spatiotemporal tracking. First, the Egg-Seg model was developed by incorporating Multi-Scale Convolutional Attention (MSCA) modules and Dice Loss to precisely segment irregular defects and egg boundaries. Second, the Egg-Weight model was established using Standard Ellipse Fitting and LightGBM to mitigate geometric distortions and optimize mass estimation on moving conveyors. Finally, the Egg-Trace method combined conveyor motion analysis with ByteTrack, implementing a “Time-Cage” mapping logic for counting and traceability. Experimental results demonstrate that Egg-Seg achieved an mAP50–95 of 93.5% for multi-class targets. The Egg-Weight model attained an R2 of 0.966 and an MAE of 0.83 g, validating robustness of the proposed feature engineering. Moreover, the system achieved 100% Accuracy of Egg Counting (AEC), 99.16% Accuracy of Egg Traceback (AET), and 97.32% Accuracy of Egg Production in Cage (AEPC), without mechanical resetting. This framework successfully digitizes multi-dimensional production metrics (quality, weight, count, location) at the individual cage level, providing important technical support for the advancement of precision livestock farming.
Unmanned aerial vehicle (UAV)-based photogrammetry provides a practical way to reconstruct farmland terrain for digital agricultural simulation, but the resulting triangular meshes preserve high-frequency microrelief and local depressions that differ fundamentally from the smooth and structured surfaces assumed in many generic vehicle simulators. When single-point wheel–ground contact is applied directly to such meshes, unstable support normals and normal forces can induce nonphysical artifacts, including geometry-induced stuck states and subsequent burst events, during low-speed operation. To address this problem, this study proposes a wheel simulation framework for rough farmland meshes based on principal component analysis (PCA)-based patch estimation (PPE). PPE aggregates multiple local contact candidates, estimates an effective wheel-patch support plane and support normal through PCA, and determines a representative contact position using soft gating. The estimated support is then coupled with an effective rigid-mesh normal-compression relation, a simplified Pacejka wheel force mapping, and explicit wheel rotational dynamics. The framework was evaluated on UAV-reconstructed farmland meshes through static support, multi-speed straight-line, and constant-steering turning tests, together with terrain-response, cross-region robustness, and computational performance analyses. Compared with single-point detection and an engine-provided wheel model, PPE consistently produced more stable support normals and normal forces, reproduced measured straight-line behavior more closely across the tested low-speed range, suppressed abnormal stuck-burst events, and preserved a steering behavior trend closer to that of a real tractor. The method also remained stable under controlled increases in mesh roughness and across real UAV-reconstructed regions with different spatial roughness patterns. Compared with flat terrain, the UAV-reconstructed terrain produced a speed-response spectrum closer to the measured tractor response and stronger simulator-level coherence between the contact-derived terrain-response indicator and vehicle speed fluctuation. Execution time profiling further confirmed that PPE satisfied the 50 Hz real-time simulation budget on the tested hardware. These results show that stable virtual tractor operation can be achieved directly on rough reconstructed farmland meshes while preserving terrain-induced response patterns relevant to real vehicle behavior, thereby supporting digital tractor simulation based on surveyed field geometry and future agricultural digital twin applications.
Air-suction seeders are essential for precision sowing of small seeds, where the accuracy and stability of the seed pickup process are crucial. To achieve high-precision prediction of seed pickup performance and interpret the influence mechanisms of operating conditions, this study proposes a prediction model and quantitative analysis method for seed pickup qualification rate based on an established multi-physics coupled model. A Uniformity-Oriented Latin Hypercube Sampling method is proposed to generate a uniformly distributed and representative training dataset, derived from a multi-physics coupled model accounting for airflow-particle-vibration in air-suction seeders during field operations. A prediction model for seed pickup qualification rate is developed using the Stacking strategy that integrates multiple elementary learners and a meta-learner. A Lévy flight-improved Black-winged Kite Algorithm (BKA-Lévy) is introduced for model optimization, enabling accurate predictions under varying disc speeds, vacuum magnitudes, and vibration frequencies and amplitudes. The model achieves R2 of 0.9215, with RMSE and MAE values of 0.0884 and 0.0715, respectively. Bench test results yield an R2 of 0.9699, further supporting the feasibility of proposed prediction model under the tested bench conditions. Finally, a surrogate model based on Gaussian Process Regression is further developed for explicit parameter-performance mapping. Kernel SHapley Additive exPlanations (Kernel SHAP) and Partial Dependence Plot (PDP) analyzes identify seeder disc speed (30 rpm) and vacuum magnitude (1700 Pa) as model-derived transition points within the parameter ranges, with vibration parameters showing strong coupling effects.
The convergence of digital agriculture and climate-smart agriculture offers promising pathways to enhance agricultural productivity, climate resilience, and environmental sustainability. However, empirical evidence on the effectiveness, implementation contexts, and limitations of digital- climate-smart agriculture integration remains fragmented. Following PRISMA 2020 guidelines, this systematic review synthesized evidence from 2145 records retrieved from Scopus, Web of Science, ScienceDirect, PubMed, and Google Scholar. After duplicate removal and rigorous screening, 75 peer-reviewed studies published between 2010 and 2025 were included. Data were extracted independently by two reviewers (Cohen's kappa = 0.82), and study quality was assessed using a modified Mixed Methods Appraisal Tool, with sensitivity analysis confirming the robustness of findings. Strong evidence from moderate to high-quality studies indicates that digital technologies, including remote sensing, IoT sensors, artificial intelligence, and mobile advisory platforms, consistently improve farm-level decision-making and resource-use efficiency. Moderate evidence supports their contribution to improved productivity and adaptive capacity, particularly through climate monitoring, precision input management, and early warning systems. Economic analyses in several regions report profitability gains and improved resilience, although outcomes vary by context and study design. Evidence for broader environmental and socio-economic impacts remains emerging, constrained by limited longitudinal evaluations and uneven geographic representation. This review makes a novel contribution by providing a transparent, evidence-based synthesis that links specific digital technologies to climate-smart agriculture outcomes, while identifying methodological, geographic, and policy gaps through a systematic appraisal. Priority research needs include longitudinal experimental studies, integrated system-level evaluations, and policy-focused implementation research in climate-vulnerable regions. These findings provide a robust evidence base to guide research, policy, and investment toward scalable and inclusive digital climate-smart agriculture.
Grid-distributed wires in T-trellis kiwifruit orchards provide structural support for dense canopy growth and enable mechanized operations such as robotic harvesting. However, these wires also introduce potential risks of collision and damage to the robot end-effector, particularly when target fruits are partially occluded by thin and visually inconspicuous wire structures. Therefore, accurate identification of wire segments and reconstruction of wire layouts are essential for safe and efficient robotic picking strategies. In this study, an identification and reconstruction method for grid-distributed wires in dense kiwifruit canopies was presented based on Geometric Feature Grouping-connection (GFG). First, fragmented wire segments were extracted from canopy images using a YOLO11x-seg model combined with an Image Overlap-partitioning and Stitching (IOS) strategy. Then, the proposed GFG approach grouped and connected discrete wire segments through directional classification, spatial grouping, and boundary extension, enabling complete reconstruction of grid-distributed wire structures. Experimental results show that the model trained on the overlap-partitioned dataset achieved an average precision of 28.50% at an Intersection over Union (IoU) threshold of 0.5, representing a 13.50% improvement over the model trained on the original dataset. With the integration of IOS strategy, the model achieved an IoU of 21.19% and a Pixel Accuracy (PA) of 21.00%, corresponding to improvements of 9.54% and 10.00%, respectively, compared with the baseline without IOS. In addition, the proposed GFG approach achieved a reconstruction accuracy of 78.28% for grid-distributed wire structures. Overall, these findings indicate that the presented method enables the reliable identification of fragmented wire segments and reconstruction of continuous grid-distributed wire structures in dense canopy environments. This study highlights the importance of reconstructing non-target support structures for improving orchard scene understanding and providing useful geometric information for future obstacle-aware robotic harvesting.
Accurate segmentation of pluckable zones and localization of plucking points are essential for intelligent tea harvesting. However, cultivar-specific traits, such as elongated internodes and low color contrast among buds, leaves, and stems, together with tea-shoot overlap and occlusion, hinder these tasks. To address these challenges, this study defines a pluckable zone for tea shoots and proposes a lightweight two-stage segmentation–localization pipeline. Stage I: an improved DeepLabV3+ model, termed DeepLabV3 + –MSEF, integrates a MobileNetV2 backbone, strip pooling–enhanced atrous spatial pyramid pooling, efficient channel attention, and focal loss for pluckable-zone segmentation. Stage II: two-dimensional (2D) plucking points are obtained from segmentation masks using a minimum bounding rectangle algorithm, and density-based spatial clustering of applications with noise (DBSCAN) is applied to the aligned three-dimensional (3D) point clouds to localize plucking points amid shoot overlap and occlusion. On an eight-cultivar test set, DeepLabV3 + -MSEF features 6.04 M parameters, runs at 60.14 frames per second (FPS), and achieves a mean intersection over union (mIoU) of 81.77% and an F1-score of 89.12%. Compared with the original DeepLabV3+ with an Xception backbone, it improves mIoU by 10.78 percentage points, reduces the parameter count from 54.71 M to 6.04 M, and nearly doubles inference speed from 30.36 FPS to 60.14 FPS, while achieving a favorable accuracy–efficiency tradeoff among benchmark models. Cultivar-wise testing demonstrates stable segmentation performance, with the mIoU ranging from 75.36% to 85.26% and F1-scores from 84.71% to 91.62%; residual errors mainly stem from variable illumination, occlusion by mature or fish leaves, and low color contrast among buds, leaves, and stems. The 2D localization stage achieves a cultivar-weighted success rate of 85.17% on 400 field images, while DBSCAN-based 3D point-cloud processing yields 92.00% plucking-point accuracy in occluded and overlapping shoot scenarios. These results demonstrate the potential of the proposed method for real-time, cross-cultivar intelligent tea harvesting.
In vertical plant factories, dynamic light management tailored to crop developmental stages is essential yet unattainable through conventional fixed-lighting approaches. This study presents an integrated control framework termed NMPC-PAR-LSTM, which synergizes three functional modules: a Michaelis-Menten kinetics model that generates growth-stage-specific photosynthetic targets, an LSTM network that forecasts environmental variations over a 20-step-ahead horizon, and an NMPC optimizer incorporating physiological constraints. Comparative evaluation demonstrates that the NMPC-PAR-LSTM system achieves a cumulative tracking error of 136 μmol·m−2·s−1, representing reductions of 91.6% relative to Linear MPC (1610.6), 90.6% relative to LSTM-MPC (1451), and 45.2% relative to MPC-PAR-LSTM (248). The framework also exhibits superior control stability, with an input variation rate of only 4.2%, compared to 26.0% for Linear MPC and 88.3% for LSTM-MPC. By aligning supplemental lighting with the physiological demands of crops, this approach provides a viable pathway toward sustainable vertical agriculture.
Deep learning has become a central component of precision agriculture, particularly for plant disease and pest detection. To enhance model performance and interpretability, attention mechanisms are increasingly integrated into convolutional and transformer-based architectures. However, their practical benefits, deployment trade-offs, and explanatory validity in agricultural settings remain insufficiently understood. This scoping review systematically examines attention-based deep learning approaches for plant health monitoring. Following PRISMA-ScR guidelines, we screened open-access studies indexed in Scopus between 2019 and 2026, resulting in 99 eligible articles. We synthesized evidence on attention mechanism design, architectural integration, optimization strategies, datasets, evaluation protocols, and explainability practices. Beyond descriptive mapping, the review provides comparative analysis of attention types under challenging field conditions and resource constraints, examines why certain mechanisms dominate current research, and clarifies their functional relevance to the visual characteristics of plant disease imagery. We further evaluate the reliability of attention-based explanations, highlighting known limitations of saliency methods and common failure modes such as attention misalignment in complex agricultural environments. In addition, we analyze how crop coverage and dataset concentration shape reported performance and generalization. Overall, attention mechanisms frequently improve predictive performance, but their benefits are strongly conditioned by data characteristics, deployment constraints, and evaluation methodology. Interpretability remains largely qualitative, and standardized comparative benchmarks are limited. This review provides a critical synthesis of methodological practices, performance trade-offs, and data dependencies, offering guidance for developing robust, interpretable, and field-deployable attention-based models for plant health monitoring.
Plant diagnostics plays a key role in optimizing plant nutrition and improving overall crop yield. For a reliable plant diagnosis, it is important to analyse both cell structure and pigment level changes to understand the stress type and the mechanisms involved. However, most of the existing techniques are unable to simultaneously visualize both changes, and some of the techniques are destructive in nature and lack sensitivity. High-resolution hyperspectral microscopy offers the capability to analyse the spectral characteristics of biological tissues, yet its application in plant diagnostics remains limited, particularly for probing deeper leaf layers. In addition, accurate characterization of different leaf layers requires high optical sectioning capability, which is lacking in conventional hyperspectral microscopes. In this context, a Structured Illumination Hyperspectral Microscope (SIHM) is developed, combining optical sectioning and hyperspectral imaging capabilities to enable layer-specific structural and spectral characterization of intact plant tissues. Sinusoidally modulated structured illumination patterns are projected onto the region of interest on the sample, and spectral images are captured using a hyperspectral imager. The developed microscope offers a spectral resolution of 1.3 nm, an optical sectioning ability of ∼5.2 μm, and a spatial resolution of ∼586 nm. The system's ability for non-destructive plant diagnostics is demonstrated through the characterization of pigment changes associated with chlorosis and necrosis, highlighting its potential for the assessment of plant health and physiology.
Modern agriculture is under increasing pressure to decarbonize, and replacing diesel tractors with their electric counterparts has gained traction as a promising pathway. The wider adoption of electric tractors faces the challenge of lacking infrastructure solutions for high-power electric tractors operating across vast, seasonal farmlands. The battery swapping stations equipped with standardized and “Lego-like” battery cells meet the high-volume charging demands and reduce downtime compared to conventional charging. To support the future infrastructure requirements of heavy-duty electric tractors, this paper addresses a critical gap in agricultural electrification by proposing a location selection and capacity configuration model for battery swapping stations and demonstrates its feasibility in a real-world scenario. By analyzing the spatiotemporal distribution pattern of the charging demand across multi-regional farmlands under a fully electrified scenario, the optimal deployment location for installing the battery swapping station is determined. Then, the capacity configuration model optimizes the number of charging bays and battery cells with the objective of minimizing the annualized investment expenditure. The case study validates the effectiveness of the deployment location selection and capacity configuration models.
Plant architecture is a pivotal agronomic trait governing cotton yield, quality, and adaptability, but its genetic dissection is hindered by the limitations of traditional manual phenotyping. To address this gap, we developed a cotton-specific integrated 3D phenomic platform based on SfM-MVS technology, equipped with two custom software tools (image acquisition & 3D reconstruction/trait extraction), which enables low-cost, high-throughput, and precise quantification of cotton architectural traits. A total of 119 genetically diverse upland cotton accessions covering nine distinct plant architecture types were subjected to 3D reconstruction. Nine machine learning models were employed to optimize architectural trait extraction, and the Bagging Decision Tree (BDT) showed the best performance, with a training coefficient of determination (R2) of 0.922 and a test R2 of 0.788. Genome-wide association study (GWAS) using plant architecture phenotypes and high-quality SNPs identified a candidate gene, GhPA1, localized on chromosome A10, which was significantly associated with multiple architectural traits. Functional validation via overexpression (GhPA1-OE) and RNA interference (GhPA1-RNAi) transgenic lines demonstrated that GhPA1 positively regulates cotton architecture, confirming the precision of digitally extracted plant architecture phenotypes. This study establishes the first comprehensive “3D phenomics-GWAS-functional validation” pipeline for cotton plant architecture research, providing a novel technical tool and genetic resource for cotton molecular breeding, as well as a transferable paradigm for dissecting complex plant architectures in other crops.
Accurate identification of fish feeding intensity is essential for intelligent feeding control and cost reduction in aquaculture. However, under complex aquaculture conditions, water surface reflections, glare interference, and high background noise often degrade the reliability of single-modal recognition methods. Although multimodal models can alleviate these limitations, they usually suffer from large parameter scales, cross-modal alignment difficulties, and limited feasibility for edge deployment. To address these issues, this paper proposes a Heterogeneous Multi-level Knowledge Distillation framework. A multi-stage audio-visual fusion network is constructed, incorporating a reliability-aware gating mechanism, a bidirectional residual interaction fusion backbone, and a robust adaptive decision fusion module to enhance cross-modal complementary representations under complex operating conditions. Meanwhile, a multi-level distillation strategy across the decision, feature, and relationship levels is designed to transfer class boundary information, key intermediate representations, and cross-modal collaborative structures from the teacher model. Furthermore, to suppress erroneous knowledge transfer caused by high-noise samples and modality-misaligned samples, an uncertainty-aware defense mechanism based on teacher prediction entropy is introduced. This mechanism dynamically adjusts the distillation intensity, enabling lightweight student networks to efficiently inherit deep latent knowledge from high-capacity teacher models. Experiments on the public AV-FFIA dataset demonstrate that the constructed student model achieves a recognition accuracy of 97.47% while reducing the number of parameters by approximately 60%, with performance approaching that of the high-capacity teacher model. The proposed method provides a feasible solution for lightweight deployment of multimodal feeding intensity recognition in complex aquaculture environments.
Extension educators assume a vital role in translating research into practical agricultural knowledge. Historically, they fulfilled this role through in-person meetings, field days, or phone conversations with farmers. In recent years, these channels have shifted to digital Extension platforms such as websites and social media. These platforms often lack sufficient innovation to enhance educator productivity while effectively addressing farmers’ spatiotemporal information and specific needs. In this work, the term Smart Agricultural Extension Platforms is used to refer to any technology that has emerged as an organizing concept within this digital transformation, integrating heterogeneous data modalities, artificial intelligence (AI) components, and decision-support functionalities into unified data-to-decision pipelines. Yet their effectiveness remains constrained by fragmented data standards and limited interoperability across the Agricultural Extension (AE) ecosystem. Consequently, this study undertakes a systematic literature review to synthesize prior works and develop a structured framework for assessing AE data pipelines. Following PRISMA principles, peer-reviewed publications and selected gray literature were screened across dedicated literature databases using defined inclusion and quality-control criteria. The analysis reveals four thematic factors that represent a critical stage in the AE data-to-decision framework. Additional findings underscore the need for technical interventions that streamline information dissemination, promote standardized data protocols, and enable the integration of AI-assisted, yet human-in-the-loop, decision-support systems. Accordingly, five research objectives were undertaken to classify AE data sources, analyze collection workflows, evaluate transformation practices, propose interoperability interventions, and define a future research agenda emphasizing validation and benchmarking.
Efficient irrigation is critical for improving agricultural water-use efficiency and crop productivity. Accurate estimation of crop evapotranspiration (ETc), a key parameter for scientific irrigation scheduling, remains challenging due to the complex coupling of external meteorological conditions and dynamic crop growth status. Single-modality data sources are insufficient to fully capture these interactions, thereby limiting estimation accuracy. To address this challenge, this study proposes M2E2Net, a multimodal deep learning model designed for high-precision, real-time daily maize ETc estimation. The model integrates smartphone-captured canopy RGB images as the visual modality, employing EfficientNet-B3 combined with Transformer to extract deep spatial and global features that accurately represent temporal crop growth dynamics. A tabular modality is constructed from RGB indices, canopy coverage, and publicly available meteorological parameters, encoded via grouped multilayer perceptrons (MLPs) to capture crop–environment interaction patterns. Cross-modal attention mechanisms are introduced to enhance semantic interaction between the two modalities, enabling dynamic matching of maize water demand variations. Validated against large-scale weighing lysimeter measurements, M2E2Net achieved an R2 of 0.8893, representing a 47.8% improvement over the conventional FAO 56 crop coefficient method. With an inference time of only 35–56 ms and lightweight design, the model is fully compatible with mobile deployment. Experiments further identified 09:00 as the optimal single-shot acquisition time. The proposed approach provides smallholder farmers with a low-cost, user-friendly, and mobile ETc monitoring solution without reliance on expensive equipment, offering substantial support for agricultural water conservation and precision irrigation.
Hyperspectral approaches offer a reliable strategy for predicting grain protein content (GPC) in winter wheat, supporting the cultivation and management of high-quality crops. However, conventional destructive sampling is labor-intensive, expensive, and unsuitable for timely monitoring. This study developed a novel multi-stage prediction framework for winter wheat grain quality prediction by integrating a mechanistic model with hyperspectral inversion of agronomic parameters (APs), enabling robust GPC prediction across different growth stages using easily accessible inputs. The results demonstrated that first-derivative spectral preprocessing combined with feature selection effectively enhanced feature representation while reducing input dimensionality. Among the multi-stage prediction models, the performance varied across growth stages when using APs as intermediate variables. Specifically, partial least squares regression with leaf nitrogen content at water ripe stage had the best prediction accuracy (R2 = 0.74, nMAE = 6.80%, and nRMSE = 5.12%). At the jointing and anthesis stages, optimal performance was obtained using partial least squares regression with leaf area index (R2 = 0.51, nMAE = 9.28%, nRMSE = 7.91%) and support vector machine with leaf nitrogen content (R2 = 0.59, nMAE = 8.46%, nRMSE = 7.03%), respectively. Further analysis of error sources revealed that model uncertainty exhibited clear stage-dependent characteristics, with minimal theoretical error at anthesis due to lower spectral inversion uncertainty, increased sensitivity to input noise at early growth stages, and degraded performance at late reproductive stages caused by canopy senescence-induced spectral signal attenuation. Compared with conventional hyperspectral approaches, the proposed framework improves prediction reliability while reducing dependence on complex inputs, thereby enhancing its applicability in large-scale and data-limited agricultural systems. Overall, this study provides both theoretical insights and a practical framework for multi-stage GPC prediction, offering strong support for precision management of wheat quality.
Effective irrigation planning and agricultural water management, particularly in regions with scarce meteorological data, fundamentally depend on the accurate calculation of reference evapotranspiration (ETo). Herein, a robust and interpretable ETo estimation framework combining heterogeneous integration feature selection (HIFS) and a stacking ensemble learning model was developed using daily meteorological data (1969–2019) from 32 sites across the Central Plains of China. HIFS created a consensus feature ranking by combining outputs from three embedded learners, extreme gradient boosting (XGBoost), random forest (RF), and gradient-boosted decision tree, through a fuzzy Borda aggregation approach. This method enabled the identification of four dominant input factors: relative humidity, wind speed at 2 m, sunshine duration, and maximum temperature. The stacking ensemble model, which integrated RF, XGBoost, and multilayer perceptron at the base level with linear regression as the metamodel, achieved peak performance with this optimal input combination (R2 = 0.9672, 95% confidence interval [CI]: 0.9653–0.969; root mean square error [RMSE] = 0.3175 mm/d, 95% CI: 0.3127–0.3223), outperforming all individual base models. Seasonal evaluation revealed that the model substantially improved winter ETo estimation (R2 improved from 0.6033 to 0.9046; errors decreased significantly, p < 0.001), a standard limitation of traditional models. Shapley additive explanation analysis confirmed the physically meaningful and seasonally varying roles of the key drivers. Neighbor-based transfer experiments revealed strong model transferability (R2 > 0.96) when data from neighboring sites were used for training. This study offers a data-efficient and interpretable alternative to the FAO-56 PM equation, providing crucial technological support for agrometeorological decision-making in agricultural regions with similar temperate monsoon climates. However, the applicability of the proposed approach in other zones warrants further investigation.
Identification of the big and small ends of poultry eggs plays an important role in promoting intelligent packaging, damage reduction, internal quality assurance and freshness. Therefore, it is key to create a high-precision, lightweight, and rapid identification model suitable for identifying the ends of poultry eggs. This study focused on three common types of eggs in the market. An improved version of YOLOv12, named RepVitASF12, was proposed to detect the big and small ends. The backbone and neck are optimized while maintaining standardization. In the backbone, a lightweight vision transformer (ViT) model replaces the original architecture to facilitate deployment. An attentional scale sequence fusion (ASF) structure is introduced that exhibits high recognition accuracy for subtle and difficult-to-identify features. Model fusion significantly enhances the network’s ability to extract and recognize the features of multiple target types. The experimental results demonstrated that RepVitASF12 outperformed YOLOv12 on multiple recognition tasks. Moreover, RepVitASF12 achieved a significant improvement in recognition accuracy (precision), achieving more than 91% for red-, white- and green-shelled eggs. For the big ends of green- and red-shelled eggs, the precision has improved. Ablation studies confirmed the effectiveness of the improved model. Although the FLOPs and model size increased slightly, the inference times for YOLOv12, YOLOv12 + RepVit, YOLOv12 + ASF and RepVitASF12 remained similar. The proposed RepVitASF12 is a high-precision, lightweight deployment potential, and rapid recognition model, and it represents a significant advancement in identifying the ends of poultry eggs, demonstrating its potential for application in automated poultry eggs packaging.
