Automatic Landslide Segmentation Using a Combination of Grad-CAM Visualization and K-Means Clustering Techniques

Abstract

Rapid detection and accurate mapping of landslides are crucial for damage detection and subsequent prevention of secondary damage. In this study, a deep learning-based segmentation model called CAM-K-SEG was proposed, which combined Grad-CAM visualization and K-Mean clustering methods to automatically detect landslide areas using satellite images. The methodology involved applying the CAM-K-SEG model to satellite images in the Bijou region of China and comparing its performance with that of K-Mean clustering and U-Net segmentation models. The optimum K value was determined by the elbow method to determine the effective color number. The weighted object was detected by removing small objects from the image, and the convolution process was performed with the mean Kernel method to remove noise or improve features. The performance of the CAM-K-SEG model was evaluated based on Intersection-Over-Union (IoU), the most used metric in semantic segmentation. The results demonstrated that the CAM-K-SEG model performed comparably to the U-Net model in segmenting landslide areas and could help improve the rapid detection of landslide areas after an event. Overall, the study contributed to the development of a new model for landslide image segmentation, which could more precisely and sensitively distinguish landslide regions. The CAM-K-SEG model was identified as a promising tool for automatic landslide detection and could be used in various applications that required accurate detection of landslide areas.