Shallow landslide detection from point clouds using shape features and deep learning

51-52

Artikel aus einem Buch

200211903

Shallow landslides pose a threat to the population, infrastructure, and ecology. Accurate and robust landslide mapping by remote sensing is an essential requirement for quantitative hazard and risk assessment, since landslide inventories are needed to calibrate landslide susceptibility models (Geitner et al., 2021, Zieher et al., 2016). For shallow landslide detection, multi-temporal 3D point clouds from topographic LiDAR have advantages compared against information derived from 2D imagery and GPS (Global Positioning System), which also have high observation accuracy and spatial resolution under forest covers (e.g., Mohan et al., 2020). The topographic complexity and the environmental changes in mountain landscapes pose challenges to change detection and assessment using point clouds. In this study, 3D shape features are particularly designed to describe local geometric structural information of 3D point clouds in both location and orientation. We embed these shape features into a hierarchical encoder-decoder deep learning framework (Hu et al., 2020) for semantic segmentation to classify terrain, buildings, grassland, and trees. Following an alignment based on building-to-building registration, we estimate topographic changes using both geometric coordinates (Lague et al., 2013) and shape features on terrain point clouds only. With this approach, we can detect shallow landslides more accurately and robustly, since the method estimates more differences (coordinates and features) and ignores vegetation changes (grassland and tree) in the point cloud comparison.
Keywords: Landslide hazards, change detection, 3D point cloud processing, shape feature representation, deep learning, scene understanding