[M-GI28] Data-driven geosciences

[MGI28-01] Statistical-mechanical prediction of landslide occurrence using structured geological map data

*Kazuhiro Miyazaki¹, Hiroki Mizuochi¹ (1.Geological Survey of Japan/National Institute of Advanced Industrial Science and Technology)

Keywords:geological map, landslide, Statistical-mechanical, neural network

We presents an advantage of structured geological map for landslide susceptibility in a wide and complex area. The structured geological map consists of separated three type of variables, such as formational age (age), petrological classification classes of rock (rock), and formational environment of rock (lithology). In contrast, the ordinary geological map (non-structured Geological map) is based on classification by integration of complex natures of age, rock type, and their formational environment. Three independent variables of the structured geological map can express thousands geologic bodies with a few ten values of each variable. We use neural network algorithm to obtain predicted probability of landslide occurrence as landslide susceptibility. The geological map data used were the 1:200,000 scale seamless geological map of Japan V2 (GSJ, 2022) from Geological Survey of Japan AIST, and the seamless 1:50,000 scale geological maps of "Imari" (Imai et al., 1959) and "Sasebo" (Matsui et al., 1989). In addition, DEM data from the Geospatial Information Authority of Japan (GSI) and landslide topographic maps from the National Research Institute for Earth Science and Disaster Resilience (NIED) were also used. The performances of landslide susceptibility calculations were compared using both structured and non-structured geological maps through receiver operating characteristic (ROC) and precision-recall (PR) analyses. The results show that the structured geological map is superior to non-structured geological map on 1:50,000 and 1:200,000 scales in the area of Northern Kyushu, southwest Japan. We also compare performances of landslide susceptibility among different areal extents of calculation using 1:200,000 structured geological map. Two types of susceptibility maps of northwestern Kyushu (hereafter Sasebo-Imari area) were obtained, such as one calculated for the Sasebo-Imari area, and the other is calculated for the entire northern Kyushu area and the results for the Sasebo-Imari area are extracted. If assumption of existence of universal probability distribution throughout the full areal extent is correct, results for the both calculations should be similar to each other. However, evaluations of ROC and PR analysis and susceptibility for each geological unit are significantly different. These results suggest that there is a possibility of an existence of overlooked random variables which varies across wide areal extent. In the other words, using a unified structured geological map to vary the calculation area for landslide susceptibility has the potential to detect hidden variables.