Detection and segmentation of loess landslides via satellite images: a two-phase framework

Landslides are catastrophic natural hazards that often lead to loss of life, property damage, and economic disruption. Image-based landslide investigations are crucial for determining landslide susceptibility and risk. In practice, satellite images have been widely utilized for such investigations; however, they still require significant labor and time resources. In this study, we propose an image-based two-phase data-driven framework for detecting and segmenting landslide regions using satellite images. In phase I, an object detection algorithm, Faster-RCNN, is trained to detect the landslide location within the large-scale satellite images. The bounding boxes of each landslide location are proposed and visualized. In phase II, we crop the satellite images into small images using the location information of the bounding boxes. Next, we use a boundary detection algorithm to identify the boundary information of each detected loess landslide to strengthen the segmentation performance. Finally, we improve the architecture of the segmentation U-Net by integrating additional inception blocks with dilation to enhance the landslide segmentation performance. A total of 150 local loess landslide occurrences in northern China are selected as our case study to validate the effectiveness, efficiency, and universality of the proposed two-phase framework. Segmentation of loess landslides is considered a challenging task due to the intrinsic nature of vague boundary information. The proposed framework is compared with the conventional U-Net and other recent benchmarking landslide segmentation algorithms. Computational results indicate that the proposed framework produces more accurate segmentation of loess landslides compared with the other tested benchmarking algorithms.