日本地球惑星科学連合2022年大会

講演情報

[J] 口頭発表

セッション記号 S (固体地球科学) » S-CG 固体地球科学複合領域・一般

[S-CG51] 機械学習による固体地球科学の牽引

2022年5月22日(日) 09:00 〜 10:30 102 (幕張メッセ国際会議場)

コンビーナ:久保 久彦(国立研究開発法人防災科学技術研究所)、コンビーナ:小寺 祐貴(気象庁気象研究所)、直井 誠(京都大学)、コンビーナ:矢野 恵佑(統計数理研究所)、座長:中野 優(国立研究開発法人 海洋研究開発機構)、加藤 慎也(京都大学防災研究所)、久保 久彦(国立研究開発法人防災科学技術研究所)

09:30 〜 09:45

[SCG51-03] Arrival time picking and Polarity detection using Deep learning for automatically creating a seismic catalog

*加藤 慎也1飯尾 能久2片尾 浩2、澤田 麻沙代2、冨阪 和秀2、水島 理恵2 (1.京都大学大学院理学研究科、2.京都大学防災研究所)

キーワード:走時pick、 極性判定、震源決定、深層学習

In recent years, the number of observation data obtained by the development of stationary observation networks like the Hi-Net, the deployment of original high dense observation networks like the Manten network, and aftershock observations has been increasing. However, since P and S wave arrival times picking and polarity of P-wave first motion determination are very time-consuming tasks, the workload is increasing as the number of data increases. Therefore, it is becoming more and more difficult for humans to perform arrival time picking and polarity determination for all the data. P and S wave arrival times have a great influence on the hypocenter determination and the estimation of seismic wave velocity structure, and the polarity of P-wave first motion affects the estimation of mechanism solution. Therefore, accurate P and S arrival times and the polarities of the P-wave first motion are required.

In this study, in order to automatically process many seismic waveforms obtained by aftershock observations and high dense observation networks with high accuracy, we created an arrival time picking model and a polarity determination model using deep learning.

The architecture used for arrival time picking model is the U-Net with Residual block and Attention block. The U-Net is a type of Fully Convolutional Network and used in previous studies (Woollam et al., 2019, Zhu and Beroza., 2019). In this study, we have created a new arrival time picking model by incorporating an attention block that learns where to focus on the data and a residual block that prevents degradation into U-Net, which has been used in previous research. On the other hand, the architecture for the polarity determination model is Convolutional Neural Network with residual blocks. In previous studies, CNN has been used for polarity determination (Hara et al., 2018, Uchide, 2020).

For the arrival time picking, we created a 3-component 100 Hz sampling model for stationary observation network, a 3-component 250 Hz sampling model for high dense observation network (Manten network), and a 1-component 100 Hz sampling model for only vertical motion components used in aftershock observation. For, the polarity determination, we created models corresponding to all sampling frequencies.

We determined hypocenters using arrival times by the model. After determining arrival times, we used REAL (Zhang et al., 2019) for phase association. For hypocenter determination, we used Hypomh_ps (Hirata and Matsu'ura. 1987; Kawanishi et al., 2009).

In this presentation, we will show the accuracy of the models and the results of the hypocenter determination and polarity determination.