Extraction of landslide mass by deep learning using 3D topographical information :

Significance of topographic map representation as training data

Abstract

  Publicly released high-resolution 3D topography information could be applied to topographical interpretation which enables to detect landslide-susceptible slopes. However, the burden on engineers or geologists is increasing due to heavy data volume for high-resolution terrain information. In the present study, deep learning was applied to interpretation of landslide topography as a method for efficiently analyzing high-resolution topography information. The deep learning was conducted based on four types of topographic maps such as a contour map, slope map, CS map and color enhancement CS map, and the target data was the Landslide Map in 1 : 50,000 scale published by the National Research Institute for Earth Science and Disaster Resilience. It was found that the locations of the landslide can be identified with a probability of up to 80％ and that accuracy of prediction was highest using a further improved CS map with color enhancement. The application of deep learning to 3D topographical information is effective in supporting the interpretation work and judgment of engineers/geologists and preventing oversight as well.