Machine Learning-based Geochemical Discrimination Method for Tsunami Deposits and a Simple Determination System

Abstract

 In order to establish a discrimination method for tsunami deposits, a machine learning analysis is conducted for geochemical data to determine paleo-tsunami deposits. Column samples containing tsunami deposits are collected at Noda village, Iwate prefecture, and Wakabayashi-ku, Sendai city, and the distribution of element concentrations are continuously measured. The model is trained by Multilayer perceptron using Noda samples as training data. Combination of elements and number of layers and perceptron are determined by the brute-force search method applied to the Noda samples. The results show that all event deposits determined in the Wakabayashi samples are tsunami deposits. These results indicate the possibility of highly accurate discrimination without being affected by sampling points or depositional ages, or by selecting appropriate supervised data. To combine the techniques of machine learning and geochemical discrimination, simple determination systematics are developed for tsunami deposits using supervised data and analyses of evaluation data.