Automatic landslide inventory generation using deep learning

With the rapid development of deep learning algorithms and easier access to remote sensing images, deep learning-based landslide identification using remote sensing images becomes possible. Pan-sharpening techniques are often adopted to fuse low-resolution multispectral images and high-resolution panchromatic images. This paper combines the deep learning and pan-sharpening techniques to enhance landslide identification results and compares the performance of four pan-sharpening techniques and two deep learning models. Eventually, morphological image processing is adopted to segment landslide clusters into individual landslides and form a basic landslide inventory. A case study of East Sai Kung, Hong Kong, shows that pan-sharpening techniques improve landslide identification accuracy and U-Net model with Brovey sharpening perform the best in this study.