Digital preservation and automated roof damage detection for traditional village heritage in Guangdong using UAV photogrammetry and deep learning

Abstract Traditional village cultural heritage (CH) in remote areas often lacks public attention and adequate maintenance. The complex spatial layouts and dense clusters of traditional buildings make manual damage inspection inefficient, delaying repairs and accelerating the loss of rural settlements. Statistics indicate that approximately two million villages disappeared in China between 1985 and 2025, with numerous rural CH sites worldwide facing similar challenges. Therefore, an efficient method for digital documentation and damage detection is urgently needed. Taking five traditional villages in Guangdong Province as case studies, this paper proposes a workflow integrating UAV photogrammetry and deep learning (DL). A total of 12762 visible-light images were acquired to construct high-precision spatial digital archives. Trained on a dataset of 41213 annotated instances, the YOLOv8-based model achieved an overall mean average precision (mAP50) of 0.832. Validation on ultra-high-resolution digital orthophoto maps (DOMs) and aerial videos demonstrated an average inference time of 5.77 ms/frame.