Performance Study of Landslide Detection Using Multi-Temporal SAR Images

Abstract

This study addresses one of the most commonly-asked questions in synthetic aperture radar (SAR)-based landslide detection: How the choice of datatypes affects the detection performance. In two examples, the 2018 Hokkaido landslides in Japan and the 2017 Putanpunas landslide in Taiwan, we utilize the Growing Split-Based Approach to obtain Bayesian probability maps for such a performance evaluation. Our result shows that the high-resolution, full-polarimetric data offers superior detection capability for landslides in forest areas, followed by single-polarimetric datasets of high spatial resolutions at various radar wavelengths. The medium-resolution single-polarimetric data have comparable performance if the landslide occupies a large area and occurs on bare surfaces, but the detection capability decays significantly for small landslides in forest areas. Our result also indicates that large local incidence angles may not necessarily hinder landslide detection, while areas of small local incidence angles may coincide with layover zones, making the data unusable for detection. The best area under curve value among all datatypes is 0.77, suggesting that the performance of SAR-based landslide detection is limited. The limitation may result from radar wave’s sensitivity to multiple physical factors, including changes in land cover types, local topography, surface roughness and soil moistures.