Attention Driven YOLOv9 for Landslide Mapping in Higher Himalayas Using Satellite Images

Detecting landslides accurately and promptly in remote sensing imagery is vital for disaster management and risk mitigation. This research employs deep learning algorithms to improve landslide detection, specifically utilizing the You Only Look Once (YOLO) framework and its variations. To enhance model performance, we incorporate three attention mechanisms—the Convolutional Block Attention Module (CBAM), the Efficient Channel Attention (ECA) module, and the global attention mechanism (GAM)—into the YOLOv9 architecture. The Nepal landslide dataset serves as our experimental dataset. We evaluate the models using metrics such as precision, recall, f-score, and mean average precision (mAP). Our findings show that the f -score for YOLOv9+CBAM is best $\mathbf{(7 9. 8 \%}$). The results highlight the efficiency and accuracy of the proposed method in landslide detection. The study underscores the potential of deep learning for disaster management and illustrates how attention modules can enhance object detection algorithms in geospatial applications.