Machine Learning Assisted Analysis on Space-Borne Point Clouds for Detecting Landslide Affected Areas

Abstract

Landslide is one of the most frequently occurred and destructive natural hazards in Taiwan and many other places around the world. Using satellite images to help identify landslide affected regions can be an effective and economic alternative comparing to conventional ground-based measures. Our previous study developed a deep learning model to analyze bi-temporal satellite images for detecting landslide affected areas in mountainous areas. The deep learning model can successfully detect spatial (planar) changes of landslides from multi-temporal satellite images. However, in a long-term monitoring of landslide affected areas, it is common to observe existing landslides occurring repeatedly. In addition to planar expansions of existing landslides and increase the extents of landslide scars, it is also common that existing landslides collapse further and produce deeper craters. Therefore, it is necessary to detect and identify the volumetric changes of landslides for a better inventory. To address this issue, this research developed a systematic machine learning framework to analyze multi-temporal three-dimensional point clouds generated from stereo pairs of high-resolution satellite images. However, the lack of appropriate and adequate training data posed a great challenge. This study first modified existing high-resolution point clouds benchmark datasets to be more consistent with the relatively low-density space-borne point clouds for preliminary training. In addition, an integration of historical LiDAR point clouds and archived satellite images were also used to generate local training datasets for transfer learning. Experimental results indicate that the developed machine learning algorithms can be used to effectively analyze space-borne point clouds for detecting volumetric changes of landslides. The results not only can produce more accurate three-dimensional landslide inventories; they are also critical factors for hazard mitigation and policy decision support.