Abstract
Pine wood nematode disease (PWD) is one of the most devastating forest diseases worldwide, often described as the "cancer" of pine trees due to its rapid and large-scale lethality. Early and accurate detection of infected trees is essential for interrupting the transmission cycle and mitigating the risk of further spread. However, current monitoring methods suffer from limited efficiency and insufficient precision. To address these challenges, this study introduces PWD-YOLO-D, an intelligent detection model for PWD based on unmanned aerial vehicle (UAV) remote sensing imagery and the YOLOv8 deep learning framework. The proposed model integrates an Efficient Multi-scale Cross-Attention (EMCA) mechanism to enhance feature representation across multiple scales and heterogeneous backgrounds; incorporates a Self-Ensemble Attention Module (SEAM) as the detection head to improve robustness in identifying occluded and overlapping diseased crowns; and adopts the Focaler-IoU loss function to refine localization accuracy and improve discrimination of complex samples. Experimental results indicate that the improved PWD-YOLO-D model outperforms the original YOLOv8 by 4.0% points in AP@0.5 and 7.3% points in AP@0.5:0.95, while reducing the Parameters by 0.48 MB. These enhancements provide strong technical support and data-driven evidence for the timely detection and precise management of infected pine trees.