Machine Learning-based Landslide Susceptibility Assessment UsingDEM-derived Topotraphic Attributes and LiDAR-derived Vegetation Attributes

Abstract

In order to develop a landslide susceptibility assessment technique, machine learning-based analysis was performed to DEM-derived topographic attributs and LiDAR-derived vegetation attributes in Hiroshima city where was severly damage by the landslides and debris flows in August 2014. We used Flow accumulation, Stream power index (SPI) and Topographic position index as topographic attributes, and tree density (N) and mean tree height (H) as vegetation attributes. From the Random forest (RF)-based analysis, we revealed that the result by the combination of SPI with N and H showed the best accuracy for assessing the landslide areas. We also confirmed that the result of the RF-based landslide susceptibility mapping showed good agreement with the distribution of the landslides in the 2014 event. The landslide susceptibility was demonstrated in other areas out of Hiroshima city affected by the 2018 debris flow event.