Multi-scale attention-enhanced deep learning approach for detecting seven trunk pests and diseases in Shanghai’s urban plane trees

Urban street trees, particularly plane trees in Shanghai, are susceptible to pests and diseases like holes, decay, termite, and longicorn. Traditional manual inspections are labor-intensive and error-prone. This study introduces an enhanced YOLOv8-based detection framework to address multi-scale variability in pest and disease datasets. Key improvements include integrating the Selective Kernel Attention (SKA) mechanism for multi-scale feature enhancement, augmenting the Spatial Pyramid Pooling Fast (SPPF) module with average pooling to preserve subtle details, and employing the WIoUv3 loss function to improve localization accuracy. Trained on 3,983 annotated samples from Shanghai, the model achieved a 3.8% increase in mean Average Precision at a 50% Intersection over Union threshold (mAP50) and a significant reduction in missed detections compared to the baseline YOLOv8. Robustness was validated through experiments with varied parameters, repeated training sessions, and stratified sampling addressing class imbalance. This research demonstrates the model's potential for intelligent, automated monitoring of urban tree health.