Japan Geoscience Union Meeting 2021

Presentation information

[J] Poster

S (Solid Earth Sciences ) » S-CG Complex & General

[S-CG52] Driving Solid Earth Science through Machine Learning

Thu. Jun 3, 2021 5:15 PM - 6:30 PM Ch.14

convener:Hisahiko Kubo(National Research Institute for Earth Science and Disaster Resilience), Yuki Kodera(Meteorological Research Institute, Japan Meteorological Agency), Makoto Naoi(Kyoto University), Keisuke Yano(The Institute of Statistical Mathematics)

5:15 PM - 6:30 PM

[SCG52-P04] Hierarchical cluster analysis of F-net moment-tensor-solution catalogue in Japan

*Hisahiko Kubo1, Takeshi Kimura1, Katsuhiko Shiomi1 (1.National Research Institute for Earth Science and Disaster Resilience)

Keywords:Hierarchical clustering, Moment tensor solution, Unsupervised machine learning, Regional characteristics of seismic activity

The spatial distribution of earthquakes is heterogeneous in response to background stresses and underground structures. Therefore, finding regional characteristics of seismic activity from seismic catalogs leads to the understanding of background stresses and underground structures that affect the way earthquakes occur. Kubo et al. (2020, SSJ) applied a dimensionality reduction technique of unsupervised machine learning to the moment tensor solution catalog of F-net of NIED and demonstrated that earthquakes with similar properties are embedded in the same region in the dimensionality-reduced space. In this study, we apply hierarchical clustering, which is also a kind of unsupervised machine learning, to the F-net moment tensor solution catalog to obtain the hierarchical structure of earthquake clusters and investigate the regional characteristics of seismic activity in Japan.

We used the moment tensor solution dataset provided by Kubo et al. (2020, SSJ). This dataset is based on the F-net moment tensor solution catalog and includes additional information on double couple and CLVD. The target parameters in the clustering of this study are latitude, longitude, and depth of earthquakes, and x and y of the source-mechanism diagram (Kavernia et al. 1996; Kagan 2005). A large x means that the reverse fault component is dominant, and a small x means that the normal fault component is dominant. A large y means that the strike-slip fault component is dominant. As a preprocessing for the cluster analysis, an axis transformation using principal component analysis was applied to latitude and longitude, and then each parameter was normalized.

There are two types of clustering methods: hierarchical clustering and non-hierarchical clustering such as k-means. Although hierarchical clustering is more computationally expensive, it has the advantage of obtaining a hierarchical structure of data with information about the relationships among clusters. In this study, we used the agglomerative method of hierarchical clustering, and the Ward method was used to calculate the distance between clusters. Agglomerative Clustering of scikit-learn was used as the analysis code.

Here, as a preliminary analysis, we examined the top 10 clusters obtained by hierarchical clustering. We found that the top 10 clusters consisted of earthquake groups with different fault types (normal fault, reverse fault, and strike-slip fault), different regions (eastern and western Japan), and different source depths (less than 20 km, 20-100 km, and greater than 100 km). This result indicates that groups of earthquakes with similar properties can be extracted by the hierarchical clustering of the moment tensor solution catalog. Further investigation of the lower clusters is expected to extract more local earthquake groups.