An Enhanced YOLO Framework for Multi-Scale Landslide Identification Under Complex Backgrounds

Deep learning has significantly improved landslide identification from remote sensing imagery, but accurately detecting multi-scale landslides under complex backgrounds remains challenging. This study proposes a lightweight YOLOv8-based model, namely YOLO-BEG, incorporating three improvements: a bidirectional feature pyramid network (BiFPN) for enhanced multi-scale feature fusion, an embedded Gaussian attention system (EGS) to improve discrimination under complex backgrounds, and a generalized intersection over union (GIoU) loss to optimize boundary localization. The model was evaluated on two datasets: a vegetation-covered Southwest landslide database and the Sichuan–Tibet Highway database. On the Southwest database, YOLO-BEG improved Precision, Recall, and F1-score by 16%, 13%, and 15% compared with YOLOv8, while using only one tenth of the parameters of Mask R-CNN. In the Sichuan–Tibet Highway database, which has more diverse background conditions, YOLO-BEG outperformed Mask R-CNN and Faster R-CNN by 32% and 13% in F1-score, respectively. These results demonstrate that YOLO-BEG is able to operate with fewer parameters and yield high-precision identification of landslides with different scales under complex backgrounds, making it a rapid and accurate tool for landslide identification.