Abstract
Sirex noctilio is an invasive pest that contributes to pine tree decline, leading to visual symptoms such as needle discoloration, crown thinning, and eventual tree death. Detecting these visible phenotypic signs from drone imagery is challenging due to elongated or irregular crown shapes, weak color differences, and occlusion within dense forests. This study introduces YOLO-PTHD, a lightweight deep learning model designed for detecting visible signs of pine decline in UAV images. The model integrates three customized components: Strip-based convolution to capture elongated tree structures, Channel-Aware Attention to enhance weak visual cues, and a scale-sensitive dynamic loss function to improve detection of minority classes and small targets. A UAV-based dataset, the Sirex Woodwasp dataset, was constructed with annotated images of weakened, and dead pine trees. YOLO-PTHD achieved an mAP of 0.923 and an F1-score of 0.866 on this dataset. To evaluate the model's generalization capability, it was further tested on the Real Pine Wilt Disease dataset from South Korea. Despite differences in tree symptoms and imaging conditions, the model maintained strong performance, demonstrating its robustness across different forest health scenarios. Field investigations targeting Sirex woodwasp in outbreak areas confirmed that the model could reliably detect damaged trees in real-world forest environments. This work demonstrates the potential of UAV-based visual analysis for large-scale phenotypic surveillance of pine health in forest management.