Analysis of Multidimensional Geochemical Data of Tsunami deposits Using Dimensionality reduction

Abstract

In this study, we proposed a machine learning method that can analyze a dataset of river sediments covering geological chemical composition data from all over Japan and tsunami deposit samples that have been accumulated so far, and analyzed them using specific samples. The samples used in the analysis were a sample identified as a muddy tsunami deposit, a sample identified as a sandy tsunami deposit, and a river deposit sample treated as a terrestrial deposit in the eastern and central Japan. The data were output as two-dimensional scatter plots by using PCA and UMAP, which are dimensionality reduction methods. The data were qualitatively evaluated by looking at the arrangement of the clusters and which of the clusters in the data set was closest to the tsunami deposit data. The results show that in the strata that contain a single layer of sandy tsunami deposits and no flood deposits, the dimensional compression of the compositional data clearly shows deposits with compositions similar to those of sandy tsunami deposits.