An Improved YOLOv10-Based Real-Time Detection Method for Apple Leaf Diseases in Complex Environments

Efficient and accurate detection of apple leaf diseases is important for timely field management and yield protection. However, real orchard images often contain complex backgrounds, uneven illumination, and small, scattered lesions, which still challenge standard one-stage detectors. In this study, we propose an enhanced real-time detector built on YOLOv10 to improve robustness and small-lesion sensitivity. First, Coordinate Attention is inserted into the backbone to encode long-range dependency while preserving precise positional cues, thereby suppressing background interference. Second, a Bidirectional Feature Pyramid Network is adopted to strengthen multi-scale feature fusion and to better retain shallow details for small targets. Third, the SIoU regression loss is introduced to accelerate convergence and improve bounding-box localization by considering distance, angle, and shape consistency. Experiments on an apple leaf disease dataset with complex field scenes show that the proposed method achieves 84.2% mAP@0.5, improving the baseline YOLOv10 by 3.5 percentage points. The detector runs at 85 FPS on a single GPU and remains suitable for real-time monitoring on edge or mobile devices. Overall, the proposed approach provides a practical and accurate solution for apple leaf disease detection in complex environments.