[S-CG45] Science of slow-to-fast earthquakes

[SCG45-13] Mosaicking Crustal Surface Morphostructure and Seismic Cycle Deformation Patterns in Chile and Japan using GNSS Velocities and Machine Learning

*Francisco H Ortega-Culaciati¹、Marcos Moreno²、Hiroki Kawabata³、Shoichi Yoshioka^4,3、Vicente Yañez-Cuadra⁵、Felipe Donoso⁵、Juan Carlos Báez⁶、Andrés Tassara⁷、Joaquin Hormazabal¹ (1.Departamento de Geofísica, Facultad de Ciencias Físicas y Matemáticas, Universidad de Chile, Santiago, Chile.、2.Departamento de Ingeniería Estructural y Geotécnica, Pontificia Universidad Católica de Chile, Santiago, Chile.、3.Department of Planetology, Graduate School of Science, Kobe University, Rokkodai-cho 1-1, Nada Ward, Kobe 657-8501, Japan.、4.Research Center for Urban Safety and Security, Kobe University, Rokkodai-cho 1-1, Nada Ward, Kobe, 657-8501, Japan.、5.Departamento de Geofísica, Facultad de Ciencias Físicas y Matemáticas, Universidad de Concepción, Concepción, Chile.、6.Centro Sismológico Nacional, Facultad de Ciencias Físicas y Matemáticas, Universidad de Chile, Santiago, Chile.、7.Departamento de Geología, Facultad de Ciencias Químicas, Universidad de Concepción, Concepción, Chile)

キーワード：clustering, machine learning, GNSS, strain, seismic cycle

We use unsupervised machine learning as an exploratory tool to analyze regional-scale patterns of crustal motions in areas affected by the seismic cycle of large subduction zone earthquakes. Particularly, we use the agglomerative clustering algorithm to investigate spatial patterns in GNSS regional velocities without the complexity of modeling a physical source. We compute surface strain and rotation rates from GNSS velocities to use them as the features needed for the clustering algorithm. Here, strain and rotation rates are estimated using a novel methodology that accounts for the spatial heterogeneity of the GNSS-network as well as for the complexities of the surface velocity field. For such purpose, we use the evidence as a Bayesian model selection method to determine the representative spatial scale at which these features are estimated. We apply the proposed methodology to analyze GNSS secular velocities in the Chilean subduction, as well as to yearly GNSS velocities over the Japanese Islands. In the Chilean subduction, the heterogeneous distribution of stations allow to identify heterogeneities in strain and rotation rates at spatial scales between 50 and 600 km, being particularly notorious the main features of regional deformation at scales > 100 km. For the Japanese Islands, the overall denser network results in the estimation of deformation features at scales between 30 to 300 km. Interestingly, our results show a spatial correlation between seismic segmentation in the fore-arc and geological structural domains. Our results demonstrate the ability of the combination of inverse and machine learning methods to efficiently identify active deformation patterns and their relationship to the subduction seismic cycle and regional-scale geological structures.