A Novel Multi-Layer Attention Boosted YOLOv10 Network for Landslide Mapping Using Remote Sensing Data

Detecting landslides is a critical challenge within the remote sensing fraternity, especially given the need for timely and accurate hazard assessment. Traditional methods for identifying landslides from remote sensing data are often manual or partially automated; however, with the progress of computer vision technology, the automated methods based on deep learning algorithms have gained significant attention. Furthermore, attention mechanisms, inspired by human visual structure, have grown remarkably in various applications, including hazard studies. In this study, we leverage the capabilities of YOLO models, especially YOLOv10 and its variants, to automate the detection of landslides. We applied four prevailing attention mechanisms: CBAM, ECA, GAM, and SA. Models are trained using the Bijie landslide detection database. Moreover, the best results are unveiled based on the evaluation criteria, that is, precision, recall, f-score, and mAP. The YOLOv10m+CBAM showed the best performance with map@50-95 of 78.5%. Our results demonstrate a robust system capable of rapidly identifying and localizing landslide events with significant detection speed and accuracy improvements. This advancement augments the process of landslide detection and supports more effective disaster response and management.