Deep generative model of landslide mass using spatial non-uniformity and continuity of landslides

Abstract

Landslide map is a thematic map used for disaster management. In recent years, there have been attempts to create landslide maps using artificial intelligence-based approaches. This study aimed to clarify the effectiveness of two normalization methods, derived from the spatial non-uniformity and continuity of landslide topography, for the deep generative model of landslide moving mass. We propose a normalization method for the supervised data to correct the spatial non-uniformity of landslides. The resulting supervised data, normalized by landslide area occupancy, improved the learning efficiency of the deep generative model. We also propose a normalization method for the inferenced results using the spatial continuity of landslides. The inferenced results, post-processed by employing our normalization method, showed reasonable distribution in comparison to the ground truth.