9:00 AM - 10:30 AM
[MGI29-P01] Comprehensive Analysis of Multidimensional Geochemical Data of Tsunami Deposits in Eastern and Central Japan Using Machine Learning
Keywords:Tsunami deposits, Machine learning, dimensional compression
Tsunami hazard map has mainly focused on the estimation of tsunami inundation areas by numerical simulation, but it has not taken into account the actual historical extent of inundation. Therefore, the purpose of this study is to establish the past tsunami inundation areas and to elaborate the tsunami hazard areas by characterizing the past tsunami deposits.
Muddy tsunami deposits, which account for 90% of the tsunami inundation area, serve as an indicator of the past tsunami inundation area. However, the conventional indicators used to discriminate tsunami deposits, such as grain size and marine microorganisms, are effective for sandy tsunami deposits, but difficult for muddy tsunami deposits. Therefore, methods based on chemical composition and data analysis have been proposed. For example, methods to determine key elements of tsunami deposits by supervised learning such as SVM (Kuwatani et al., (2014)) and unsupervised learning such as hierarchical clustering to determine tsunami deposit indices (Watanabe et al., (2021)). However, these previous studies have focused on specific areas, mainly in the Tohoku region, where samples were collected, and have also focused on specific elements. Therefore, it is not sufficient to elucidate the collective behavior of many elements in muddy tsunami deposits over a wide area and to propose a discrimination method. In addition, it is necessary to be able to analyze tsunami deposits in western Japan in order to prepare for future Nankai Trough earthquakes.
In this study, we attempted to elucidate the geochemistry of tsunami deposits in western Japan by using the geochemical map of Japan (AIST, 2010) and a machine learning method that can comprehensively analyze tsunami deposit samples accumulated up to now. The samples included terrestrial sediment data and ocean bottom sediment data from the Chubu, Kanto, and Tohoku regions in the geochemical map, muddy tsunami sediment data from the Great East Japan Earthquake, and soil samples from the coastal areas of Noda Village, Iwate Prefecture (Noda Village data) and Suruga Bay, Shizuoka City, Shizuoka Prefecture (Shizuoka data) where multiple tsunami deposits were identified. Borehole core samples were handled. Concentration data of 17 elements consisting of these major elements and some heavy metal elements, etc., were output to two-dimensional scatter plots reflecting complex concentration distribution patterns by using dimensional compression methods, PCA and UMAP, and were qualitatively evaluated.
As a result, UMAP results showed strong similarity between ocean bottom sediment and land sediment data via muddy tsunami sediment data. This indicates that the chemical analysis of muddy tsunami sediments is likely to include a mixture of samples with strong similarity to terrestrial sediments. Therefore, the results suggest the need to carefully select data that do not resemble terrestrial sediments when dealing with teacher data of muddy tsunami deposits for tsunami sediment data analysis. In addition, a cluster was generated in UMAP that included data already identified as tsunami deposits, muddy tsunami deposit data, and ocean bottom deposit data for Noda-mura data, and some Noda-mura data that were not identified as tsunami deposits were included in this cluster. This indicates that we were able to identify data with a high likelihood of being tsunami deposits. Since all Shizuoka data were generated within the terrestrial sediment cluster, it is clear that a new element that can account for the hinterland needs to be introduced.
Muddy tsunami deposits, which account for 90% of the tsunami inundation area, serve as an indicator of the past tsunami inundation area. However, the conventional indicators used to discriminate tsunami deposits, such as grain size and marine microorganisms, are effective for sandy tsunami deposits, but difficult for muddy tsunami deposits. Therefore, methods based on chemical composition and data analysis have been proposed. For example, methods to determine key elements of tsunami deposits by supervised learning such as SVM (Kuwatani et al., (2014)) and unsupervised learning such as hierarchical clustering to determine tsunami deposit indices (Watanabe et al., (2021)). However, these previous studies have focused on specific areas, mainly in the Tohoku region, where samples were collected, and have also focused on specific elements. Therefore, it is not sufficient to elucidate the collective behavior of many elements in muddy tsunami deposits over a wide area and to propose a discrimination method. In addition, it is necessary to be able to analyze tsunami deposits in western Japan in order to prepare for future Nankai Trough earthquakes.
In this study, we attempted to elucidate the geochemistry of tsunami deposits in western Japan by using the geochemical map of Japan (AIST, 2010) and a machine learning method that can comprehensively analyze tsunami deposit samples accumulated up to now. The samples included terrestrial sediment data and ocean bottom sediment data from the Chubu, Kanto, and Tohoku regions in the geochemical map, muddy tsunami sediment data from the Great East Japan Earthquake, and soil samples from the coastal areas of Noda Village, Iwate Prefecture (Noda Village data) and Suruga Bay, Shizuoka City, Shizuoka Prefecture (Shizuoka data) where multiple tsunami deposits were identified. Borehole core samples were handled. Concentration data of 17 elements consisting of these major elements and some heavy metal elements, etc., were output to two-dimensional scatter plots reflecting complex concentration distribution patterns by using dimensional compression methods, PCA and UMAP, and were qualitatively evaluated.
As a result, UMAP results showed strong similarity between ocean bottom sediment and land sediment data via muddy tsunami sediment data. This indicates that the chemical analysis of muddy tsunami sediments is likely to include a mixture of samples with strong similarity to terrestrial sediments. Therefore, the results suggest the need to carefully select data that do not resemble terrestrial sediments when dealing with teacher data of muddy tsunami deposits for tsunami sediment data analysis. In addition, a cluster was generated in UMAP that included data already identified as tsunami deposits, muddy tsunami deposit data, and ocean bottom deposit data for Noda-mura data, and some Noda-mura data that were not identified as tsunami deposits were included in this cluster. This indicates that we were able to identify data with a high likelihood of being tsunami deposits. Since all Shizuoka data were generated within the terrestrial sediment cluster, it is clear that a new element that can account for the hinterland needs to be introduced.