Qinghai-Tibet Plateau (QTP) Landslides Dataset

Optical remote sensing images, with their high spatial resolution and wide coverage, have emerged as invaluable tools for landslide analysis. Visual interpretation and manual delimitation of landslide areas in optical remote sensing images by human is labor intensive and inefficient. Automatic delimitation of landslide areas empowered by deep learning methods has drawn tremendous attention in recent years. Mask R-CNN and U-Net are the two most popular deep learning frameworks for image segmentation in computer vision. In this study, we systematically compare and evaluate the performance and adaptability of the Mask R-CNN and the U-Net models for delimiting landslide areas in optical remote sensing images with various resolutions across regions using statistical metrics. A workflow for transferring deep learning models pretrained on other regions for landslide area delimitation on new regions with a relatively small number of annotated training samples is developed. A post-processing module is integrated into the Mask R-CNN architecture to address the challenge of overlapping mask predictions for individual landslide objects. The results indicate that the Mask R-CNN model exhibits superior overall performance in comparison with the U-Net model and is more suitable for tasks requiring detailed delineation of the object outlines in images. The insights gained from this study not only advance our understanding of the models' generalizability and robustness across regions, but also have practical implications for large-scale landslide inventory mapping in remote sensing images.