Assessment for sediment disasters considering topographical features of the neighborhood by deep learning

Takeharu SATO


This paper proposes a new method for assessing the sediment disaster risk using a deep‐learning‐based analysis of Digital Elevation Models (DEMs). The following three features are involved in this method:1) conducting the risk assessment to each grid cell of DEMs, 2) calculating multiple terrain quantities related to sediment disasters, 3) incorporating values of the terrain quantities from surrounding cells for each cell. Three terrain quantities, gradient, Laplacian, and specific height difference, were used as explanatory variables of the sediment disaster risk. The author generated a data set of the explanatory variables by incorporating adjacent 7 by 7 grid cells into each cell. Based on deep learning training with this data set, a hierarchical neural network consisting of the five‐layer structure with three intermediate layers was selected as the optimal model. The study area was set in the upstream region of the Tsurugi River catchment, Hofu City, Yamaguchi Prefecture. The model appropriately detected about 85% of landslide sites induced by the 2005 Hofu heavy rain disaster. The model analysis also showed that the sediment disaster risk became higher in 0‐order basin valley, which is conventionally considered landslide‐prone areas in mountain regions.

Key words

Digital Elevation Model (DEM), terrain quantity, deep learning, risk assessment, sediment disasters