Xing Yang

A dominant theme across the collection is the development of real-time detectors for crop diseases, insect pests, small lesions, eggs, and visually defective plant products. These studies share a practical recognition that field targets are rarely large, isolated, or cleanly bounded. Improved YOLO and RT-DETR variants are accordingly applied to small-object detection and background interference in diverse crop scenes (Lv et al., 2026;Liu et al., 2026). Related studies address multi-scale fusion, boundary ambiguity, and class imbalance in disease or pest recognition tasks (Li et al., 2026;Ye et al., 2026). Further contributions extend real-time perception across maize, pakchoi, litchi, sugarcane, and other production scenes (Liu et al., 2026;Yang et al., 2026). Additional work broadens this direction to plant-product quality and crop-specific visual targets (Zhou et al., 2026;Zhu et al., 2026). Together with studies on defective products and additional crop targets, these articles show how object detection is being adapted to the visual diversity of agricultural production (Ma et al., 2025;Li et al., 2025).The same perceptual challenge recurs in specialized pest and forest-monitoring scenarios. Studies targeting small irregular foliar lesions and trap-based insect identification extend visual AI beyond standard leaf-image classification (Zeng and Tao, 2026;Li et al., 2026). Intensive small-target pest detection and Pomacea canaliculata egg-mass recognition further illustrate the pressure placed on detectors by occlusion, scale variation, and irregular target morphology (Liu et al., 2025;Wang et al., 2025). Forest and orchard monitoring introduce still broader scene complexity and operational constraints (Hu and Wang, 2025;Shen et al., 2025). The pine wilt disease study likewise emphasizes the need for scene-specific perception systems in large-scale ecological monitoring (Xiao et al., 2026). Collectively, these articles signal a clear transition from generic model improvement toward crop-, pest-, and scene-specific perception systems.Several articles address the persistent tension between model performance and deployability. Lightweight convolutional networks and attention-enhanced backbones are used to improve diagnostic accuracy while reducing computational burden (Shi et al., 2026;Adithiyaa et al., 2026). Transformer-based classifiers and hybrid models further show how architectural choices can balance accuracy, robustness, and implementation cost (Xu et al., 2025;Inamdar et al., 2026). Fusion models and co-optimization strategies extend this theme by treating efficiency as part of the diagnostic objective rather than an afterthought (Liu et al., 2025;Chang et al., 2025). Additional compact and task-adapted designs reinforce the same deployment-aware direction across field-diagnosis settings (Meenakshi and Sankaranarayanan, 2026;Ashraf et al., 2026). Related optimization of diagnostic models continues this movement toward practical inference in agricultural environments (Wu et al., 2026). This direction is particularly critical for sustainable agriculture, since a technically sophisticated model holds limited practical value if it cannot be deployed on affordable hardware or maintained within routine production workflows. The same deployment challenge is evident in Solar Insecticidal Lamp Internet of Things (SIL-IoT) systems, where resource constraints and harsh outdoor conditions require fault detection to be both reliable and lightweight (Yang et al., 2023;Yang et al., 2026a). Recent SIL-IoT visual health monitoring further links device reliability with multi-task perception and logic-guided consistency, showing that intelligent plant protection also depends on the condition-awareness of its sensing infrastructure (Yang et al., 2026b).The collection also highlights learning under limited, heterogeneous, and user-generated data. Few-shot recognition and interpretable multi-view insect classification respond to the reality that agricultural data are often unevenly distributed, region-specific, and expensive to annotate (Zhang et al., 2026;Liu et al., 2026).Soft prompt tuning based on colloquial farmer descriptions and procedural synthetic data generation further broaden learning beyond curated expert datasets (Liu et al., 2025;Farinati Leite et al., 2025). These studies broaden the field from pure image recognition to agricultural intelligence that can operate effectively with scarce samples, informal descriptions, and incomplete field evidence.Trustworthiness emerges as another important thread. One study directly quantifies the reliability gap between controlled laboratory imagery and field imagery, demonstrating why high benchmark accuracy may not translate reliably to deployment conditions (Xiang et al., 2026). Another foregrounds confidence calibration and uncertainty estimation for grape leaf disease diagnosis under real-world field conditions (Ergun et al., 2026). Together, these contributions make a compelling editorial point: intelligent plantprotection systems should report not only what they predict, but also when and why those predictions are likely to be unreliable.Several articles move beyond single RGB images by incorporating richer sensing modalities and more field-representative data. RGB-NIR fusion is utilized to enhance weak lesion and pest-damage contrast in tea plantations, while hyperspectral analysis is explored for differentiating grapevine yellows diseases from symptomatically similar abiotic stresses (Wang et al., 2026;Alisaac et al., 2026). UAV imagery further extends the diagnostic scale from individual leaves to entire crop stands and field environments (Zhang et al., 2025). These studies suggest that future progress will depend as much on sensing design and acquisition context as on network architecture.Segmentation and spatial localization form a complementary theme. Pixel-level or region-level methods applied to paddy pests and pine wilt disease...

Introduction/Objective: Camera-based pig behavior detection can support livestock monitoring, but performance claims are difficult to interpret when public image datasets are reorganized without transparent split control or when ensemble parameters are selected using the final test set. This study aimed to construct controlled public-image benchmarks for pig behavior detection and evaluate whether validation-selected decision-level fusion improves offline detection accuracy. Methods: Two public pig behavior datasets were reorganized into six-class and twelve-class benchmarks. After grouping images by original base filename, Pig-6 contained 16,608 training images, 1,779 validation images, and 1,779 test images, while Pig-12 contained 4,034 training images, 439 validation images, and 439 test images. Classical detectors, YOLO-family detectors, and available transformer detectors were evaluated with replicated seeds using a unified 101-point average-precision evaluator. Component detectors and fusion parameters were selected only on the validation split, and final predictions were combined using weighted box fusion (WBF) and test-time augmentation (TTA) before a single held-out test evaluation. Results: On Pig-6, mAP50-95 increased from 55.86% ± 0.16% for the strongest replicated single detector to 56.62% for the final WBF+TTA ensemble; image-level bootstrap analysis showed a paired gain of 0.59 percentage points over the best validation-selected single run, with a 95% confidence interval of 0.20-1.11 points. On Pig-12, mAP50-95 increased from 70.31% ± 0.84% to 73.05%, with a paired bootstrap gain of 1.89 points and a 95% confidence interval of 0.19-3.73 points. Control ablations showed that the final WBF+TTA configuration exceeded the no-TTA WBF control on Pig-6 (55.82%) and the NMS-union TTA control on Pig-12 (72.15%). Discussion: Class-wise analysis showed stronger performance for feeder, waterer, and common posture categories, whereas rare posture-drinking combinations remained difficult. The results indicate that validation-selected decision-level fusion was more beneficial for the finergrained Pig-12 benchmark, although the final ensemble also increased computational cost. Conclusion: Validation-selected WBF-TTA fusion improved performance on the two evaluated public benchmarks under the reported offline protocol, while external-farm validation, temporal behavior recognition, and deployment efficiency remain unresolved.

As an innovative application of wireless sensor networks (WSNs), the Solar Insecticidal Lamp Internet of Things (SIL-IoTs) is widely deployed for green pest control. However, existing deployment strategies commonly assume homogeneous energy-harvesting capabilities among nodes and mandate full coverage, thereby overlooking the realities of uneven solar irradiation and budget constraints in complex 3D terrains. To address these issues, this paper investigates the SIL deployment problem using real DEM data. We first introduce a 3D coverage model alongside a terrain-based solar energy weight model, and formulate the task as a partial coverage optimization problem constrained by both energy harvesting and budget limitations. To solve this problem and avoid premature convergence, we propose an Improved Jumping Particle Swarm Optimization (IJPSO) algorithm. Experimental results on benchmark datasets and DEM-based 3D terrain simulations demonstrate that IJPSO outperforms six competing algorithms, achieving average relative percentage deviations (RPD) of 0.045% and 0.140% and average 3D deployment objective values of 62.27 and 84.37 under the 80% and 90% coverage constraints, respectively. These results demonstrate the effectiveness of IJPSO in reducing deployment costs while maintaining solution consistency.

Reliable fault diagnosis of field-deployed intelligent agricultural equipment requires not only accurate recognition of known faults but also reliable rejection of previously unseen fault conditions encountered after deployment. This challenge is particularly pronounced in solar insecticidal lamp Internet of Things (SIL-IoTs) systems, where environmental variability, component aging, and operational disturbances can induce substantial intra-class variability and compromise the reliability of conventional closed-set diagnosis. To address this challenge, this article proposes a prototype-based open-set fault diagnosis framework for SIL-IoTs systems. A temporal encoder is trained to extract discriminative representations from multichannel electrical signals. Within the learned feature space, class prototypes are constructed for known faults, and an uncertainty-aware prototype distance (UAPD) is introduced to account for class-wise feature variance during prototype matching. A class-adaptive conformal thresholding (CACT) strategy is further developed to estimate an independent acceptance threshold for each known class using only calibration samples from known classes, thereby enabling open-set rejection without requiring real unknown faults for threshold tuning. Experiments on five open-set diagnosis tasks show that the proposed method achieves the highest overall open-set accuracy across all tasks and the highest H-score in four tasks, with absolute gains of up to 0.0377 over the strongest baseline. These results demonstrate that the proposed framework provides an effective and practically applicable solution for open-set fault diagnosis in field-deployed SIL-IoTs systems.

Reliable solar insecticidal lamp Internet of Things (SIL-IoT) equipment underpins green pest control and precise pest monitoring. Healthy ranges of light, photovoltaic and thermal variables shift with daylight, weather, energy availability and installation conditions, while labelled faults remain scarce and imbalanced. We introduce a diagnosability-aware operating-state health baseline (DA-OHB) for calibration-based cross-node diagnosis of data-observable faults in SIL-IoT equipment. Candidate faults were screened by data observability, mechanistic expressibility and availability as curated telemetry event labels. Photovoltaic–light, electrical-box/air-temperature, power and rolling-state features were combined with operating-state gates, direction-sensitive evidence scores and target-node healthy false-alarm calibration. We evaluated DA-OHB on July–August 2025 field records from four devices deployed in Chuzhou, China, for three maintenance-relevant faults: light-intensity sensor open circuit, light-intensity/solar-panel-current mismatch, and electrical-box/air-temperature mismatch. Under leave-one-device-out aggregation with an early healthy calibration subset from each target node, mean F1-scores were 0.996, 0.823 and 0.770; mean area under the precision–recall curve values were 1.000, 0.890 and 0.963. Across five seeds, DA-OHB F1-scores were 0.996 ± 0.000, 0.823 ± 0.000 and 0.763 ± 0.004. F1 reflects the target-node mechanism with sufficient fault evidence, whereas F2 and F3 demonstrate cross-node evidence from multiple devices. Field diagnosis of SIL-IoT equipment thus benefits from linking alarms to valid operating states, fault directions and node-specific healthy calibration. DA-OHB provides an interpretable basis for SIL-IoT maintenance under curated telemetry labels; effects on pest-count estimates and agricultural decisions require separate evaluation.

Solar insecticidal lamps are important physical control devices for green pest management, but faults in their photovoltaic power supply units can reduce trapping efficiency and shorten service life. To improve fault identification under complex agricultural environments, this study proposes a signal-image-level multimodal fusion network (SIL-MMFN) for detecting and classifying photovoltaic panel operating states in solar insecticidal lamps. The method combines time-series measurements with short-time Fourier transform (STFT)-based time–frequency images. A convolutional image branch extracts spatial features from time–frequency representations, whereas a bidirectional GRU branch with attention models temporal dependencies in the original signals. In addition, physics-informed features based on the illumination–current residual and output power are introduced to enhance discriminative fault information. Field data collected from four agricultural deployment nodes were used to classify normal, open-circuit, and mismatch states. Experimental results show that the proposed method achieved an accuracy of 97.5%, precision of 96.7%, recall of 97.8%, and macro-F1 score of 97.3%, outperforming single-modality and representative comparison models. The results indicate that multimodal fusion helps reduce confusion between open-circuit and mismatch faults and provides a potential approach for operating-state monitoring and maintenance of agricultural photovoltaic equipment. In this study, fault diagnosis refers to the detection and classification of photovoltaic panel operating states, including normal, open-circuit, and mismatch conditions.

Ensuring food security requires innovative, sustainable pest management solutions. The Solar Insecticidal Lamp Internet of Things (SIL-IoT) represents such an advancement, yet its reliability in harsh, variable outdoor environments is compromised by frequent component and sensor faults, threatening effective pest control and data integrity. This paper presents a comprehensive survey on fault detection (FD) for SIL-IoT systems, systematically analyzing their unique challenges, including electromagnetic interference, resource constraints, data scarcity, and network instability. To address these challenges, we investigate countermeasures, including blind source separation for signal decomposition under interference, lightweight model techniques for edge deployment, and transfer/self-supervised learning for low-cost fault modeling across diverse agricultural scenarios. A dedicated case study, utilizing sensor fault data of SIL-IoT, demonstrates the efficacy of these approaches: an empirical mode decomposition-enhanced model achieved 97.89% accuracy, while a depthwise separable-based convolutional neural network variant reduced computational cost by 88.7% with comparable performance. This survey not only synthesizes the state of the art but also provides a structured framework and actionable insights for developing robust, efficient, and scalable FD solutions, thereby enhancing the operational reliability and sustainability of SIL-IoT systems.