Abstract:

Pine wilt disease poses a serious threat to global pine resources and ecological environment.Accurate detection of infected trees is critical to prevent further spread of the disease.We used unmanned aerial vehicle(UAV)-based remote sensing technology for the efficient acquisition of extensive and high-resolution imagery of forested areas,providing crucial data support for the detection of PWD-infected pine trees.To address the limitation in detection capability of PWD-infected pine trees within UAV remote sensing imagery under complex forest environments,we presented an enhanced YOLOv8n detection model called YOLOv8n-RCD.The model employs RepVit as the backbone network to improve feature extraction capability,integrates a Cross-scale Convolutional Feature Fusion Module(CCFM)to strengthen multi-level feature extraction,and employs Dynamic Head in place of the original detection head,thereby enhancing target recognition and adaptability in complex backgrounds.Experimental evaluations demonstrated that the improved model of YOLOv8n-RCD achieved relative gains of 3.37%,3.00%,and 3.19% in precision(P),recall(R),and F₁ score,respectively,over the baseline model YOLOv8n,and its AP50 and AP50-95 were increased by 1.93% and 1.49% compared with the latter.The enhanced model improved detection accuracy and recognition capability in complex forest environments,providing robust technical support for precise identification and UAV remote sensing-based intelligent monitoring of PWD-infected pine trees.

Key words: object detection, pine wilt disease, YOLOv8n, RepVit, cross-scale convolutional feature fusion module, dynamic head