Attention Driven YOLOv5 Network for Enhanced Landslide Detection Using Satellite Imagery of Complex Terrain

Landslide hazard detection is a prevalent problem in remote sensing studies, particularly with the technological advancement of computer vision. With the continuous and exceptional growth of the computational environment, the manual and partially automated procedure of landslide detection from remotely sensed images has shifted toward automatic methods with deep learning. Furthermore, attention models, driven by human visual procedures, have become vital in natural hazard-related studies. Hence, this paper proposes an enhanced YOLOv5 (You Only Look Once version 5) network for improved satellite-based landslide detection, embedded with two popular attention modules: CBAM (Convolutional Block Attention Module) and ECA (Efficient Channel Attention). These attention mechanisms are incorporated into the backbone and neck of the YOLOv5 architecture, distinctly, and evaluated across three YOLOv5 variants: nano (n), small (s), and medium (m). The experiments use open-source satellite images from three distinct regions with complex terrain. The standard metrics, including F-score, precision, recall, and mean average precision (mAP), are computed for quantitative assessment. The YOLOv5n + CBAM demonstrates the most optimal results with an F-score of 77.2%, confirming its effectiveness. The suggested attention-driven architecture augments detection accuracy, supporting post-landslide event assessment and recovery.