Detecting Landslides with Deformable Adaptive Focal YOLO: Enhanced Performance with Reduced False Detection

Landslide detection algorithms employing You Only Look Once (YOLO) often exhibit subpar performance in scenarios characterized by similar backgrounds and boundary blur, leading to high miss and false rates. To address these challenges, this article presents a novel approach, namely Deformable Adaptive Focal YOLO (DAF-YOLO), which builds upon YOLOv8. First, the Enhanced Deformation Convolutional Network (EDCN) is crafted with the purpose of enhancing the discernment of anomalous landslides. Secondly, a lightweight Sliding Window Attention Mechanism (SW-ATT) is constructed to refine the capacity for discriminating backgrounds through the use of localized windows and enhanced similarity measurement techniques, alongside similarity activation maps. Lastly, an Adaptive Varifocal Loss (AVFL) framework is proposed to effectively mitigate both missed detections and false detection. Experimental results show that compared with YOLOv8, DAF-YOLO has significantly improved the F1 index and mAP, reaching 0.940 and 0.962 respectively, while maintaining a detection accuracy of 0.941 and a recall rate of 0.940.