Machine Learning Identification of Multidimensional Geochemical Data of Tsunami Deposits

Abstract

The purpose of this study is to clarify the distribution area of past tsunami deposits and to refine the tsunami hazard area. Muddy tsunami deposits, which cover 90% of the tsunami inundation area, are an indicator of the past tsunami inundation area. However, comprehensive characterization of muddy tsunami sediments over a wide area using chemical compositions and a discrimination method have not been sufficiently proposed. In this study, we evaluate the properties of tsunami deposits in the Tohoku region based on multidimensional chemical composition data sets of fluvial and multi-regional tsunami deposits throughout Japan.The data sets used were river sediments, the 2011 Tohoku-oki tsunami sediments, and analytical data from tsunami sediments in Noda-mura and Hachinohe, Iwate Prefecture, and Higashimatsushima, Miyagi Prefecture. The elements used were 29 elements and the Na/Ti ratio, one of the indicators of tsunami deposits. The analysis was performed using scatter plots and UMAP, one of the dimensional compression methods, to confirm the certainty of the estimation results obtained from the scatter plots.As a result, it is possible that tsunami deposits and land source deposits can be discriminated, at least in the Tohoku region, by referring to the Fe, Zr, and As associations in parallel.