Detecting citrus Huanglongbing (HLB) in orchard environments is particularly challenging due to multi-scale targets and occlusions due to clustering, which manifest as complex and variable backgrounds, targets ranging from distant single leaves to nearby full canopies, and frequent instances where symptomatic leaves are hidden behind others, all significantly hindering accurate detection. To overcome these challenges, this study introduces a novel citrus object detection model, Multi-Scale Huanglongbing DETR (MSHLB-DETR), developed on the basis of an improved Real-Time DEtection TRansformer (RT-DETR). The model significantly enhances detection accuracy and efficiency for HLB under complex orchard conditions. To address the issue of small target feature loss in leaf detection, a new efficient transformer module called Smart Disease Recognition for Citrus Huanglongbing with Multi-scale (SDRM) is introduced. SDRM includes a space-to-depth (SPD) module and inverted residual mobile block (IRMB), which facilitate deep interaction between local and global features and significantly improve the computational efficiency of the transformer. Additionally, the transformer encoder incorporates a Context-Guided Block (CGBlock) for contextual feature learning. To evaluate the proposed model under complex background conditions, a dataset of 4367 images was collected from diverse orchard scenes, preprocessed, and divided into training, validation, and testing subsets. The experimental results demonstrate that the proposed MSHLB-DETR achieved the best detection performance on the test set, with an ${\mathrm{mAP}}_{50}$ of 96.0%, surpassing other state-of-the-art models of similar scale. Compared to the original RT-DETR, the proposed model increased ${\mathrm{mAP}}_{50}$ by 15.8%, reduced Params by 7.5%, and decreased GFLOPs by 5.2%. This study reveals the critical importance of developing efficient multi-scale detection techniques for the accurate identification of citrus Huanglongbing in complex real-time monitoring scenarios. The proposed algorithm is expected to provide valuable references and new insights for the precise and timely detection of citrus Huanglongbing.

Effective monitoring of protected wildlife is crucial for biodiversity conservation. While camera traps provide valuable data for ecological observation, existing deep learning models often suffer from low accuracy in detecting rare species and high computational costs, hindering their deployment on edge devices. To address these challenges, this study proposes YOLO11-APS, an improved lightweight model for protected wildlife detection. It enhances the YOLO11n by integrating the self-Attention and Convolution (ACmix) module, the Partial Convolution (PConv) module, and the SlimNeck paradigm. These improvements strengthen feature extraction under complex conditions while reducing computational costs. Experimental results demonstrate that YOLO11-APS achieves superior detection performance compared to the baseline model, attaining a precision of 92.7%, a recall of 87.0%, an mAP@0.5 of 92.6% and an mAP@0.5:0.95 of 62.2%. In terms of model lightweighting, YOLO11-APS reduces the number of parameters, floating-point operations, and model size by 10.1%, 11.1%, and 9.5%, respectively. YOLO11-APS achieves an optimal balance between accuracy and model complexity, outperforming existing mainstream lightweight detection models. Furthermore, tests on unseen wildlife data confirm its strong transferability and robustness. This work provides an efficient deep learning tool for automated wildlife monitoring in protected areas, facilitating the development of intelligent ecological sensing systems.

Small pest detection in ultra-high-resolution forestry images is challenging due to extreme scale variation, complex backgrounds, and limited annotated data. To address these issues, we propose SSFPDet (Semi-Supervised Forest Pest Detector), a semi-supervised object detection framework designed for low-annotation settings. Built upon the Soft Teacher paradigm, SSFPDet integrates a YOLO-T-based overlapping slicing strategy, a Top-K pseudo-label selection mechanism, and a Kullback–Leibler (KL) divergence-based distribution alignment constraint. The slicing strategy enhances small-object representation without modifying the detector backbone, while the Top-K and KL modules improve pseudo-label reliability and semantic consistency during training. Under the 20% labeled setting, SSFPDet achieves an mAP@0.5:0.95 of 46.6, outperforming the baseline by 0.7 points. Notably, small-object detection performance (AP_S) improves by 6.6 percentage points. Ablation studies confirm the complementary contributions of spatial slicing and semantic alignment. Overall, SSFPDet provides a practical and scalable solution for high-resolution forestry pest monitoring under limited supervision.