Attention-Enhanced YOLOv8s Framework for Accurate Small-Scale Landslide Detection from Remote Sensing Imagery

Landslides are among the most destructive natural hazards in mountainous regions, posing severe risks to human life, infrastructure, and sustainable development. Rapid and accurate post-disaster landslide detection is crucial for emergency response and risk mitigation. This study presents an enhanced YOLOv8s-based detection framework integrated with the Convolutional Block Attention Module (CBAM) to improve the recognition of small-scale landslides in complex remote sensing imagery. The publicly available Bijie landslide dataset from Wuhan University was used, with multiple augmentation strategies applied to increase feature diversity and reduce overfitting. The CBAM was embedded within the YOLOv8s architecture to strengthen spatial–channel attention and enhance multi-scale feature fusion, while the Efficient Intersection over Union (EIoU) loss function refined bounding box regression. Experimental results indicate that the proposed YOLOv8s-CBAM model achieves superior performance, reaching a precision of 92.1%, recall of 86.8%, and mAP@0.5 of 89.6%, representing relative improvements of 14.9%, 3.5%, and 6.9% over the baseline YOLOv8s model, respectively. The enhanced model demonstrates strong robustness under complex illumination, vegetation, and terrain conditions. These findings suggest that YOLOv8s-CBAM provides a reliable, real-time, and scalable framework for automated landslide detection, with significant potential for operational applications in large-scale disaster monitoring and rapid emergency assessment.