11:00 AM - 1:00 PM
[SCG51-P02] Application of deep learning picker to automatically determined earthquakes around the 2011 Tohoku earthquake
Keywords:Phase picking, Hypocenter determination, Deep learning
A seismic phase picking is the base of the seismological analysis. The JMA has been used the AR-AIC method to pick P- and S-phases for the production of an earthquake catalog. The AR-AIC method sometimes reads non-seismic signals (noise) in waveforms. In recent years, deep learning phase pickers including phase classification of P- and S-wave have been proposed. In this study, we adopted deep learning pickers in the conventional automatic processing and evaluated hypocenters for multi-channel long-term continuous waveforms.
We used PhaseNet (Zhu and Beroza 2019) and EQTransformer (Mousavi et al. 2020). Both pickers were trained using global seismic waveforms to detect and classify P- and S-phases. We used the PF method (Tamaribuchi et al. 2016, 2018) for phase association and hypocenter determination. The PF method is currently applied to the automatic hypocenter determination of the JMA catalog.
To evaluate the performance in the normal and swarm period, we used the continuous waveforms at about 1300 stations nationwide for 12 hours each from 0:00 on March 1, 2011 and 12:00 on March 11, 2011. Since it takes a too long time to read the waveform data in WIN format directly with python, we converted it to mini SEED format and then performed phase pickers. As a result, it took about 2 hours and 30 minutes in the total workflow including the format conversion, picking, and hypocenter determination processing for the 12-hour waveforms.
As for the result of EQTransformer, we determined 278 events automatically, compared to 196 events in the JMA catalog for 12 hours from 0:00 on March 1, 2011. We also determined 862 events automatically, compared to 1192 events in the JMA catalog for 12 hours from 12:00 on March 11, 2011. Although the number of noises (false-positive) has decreased compared to the conventional PF method, the number of missing earthquakes (false-negative) has increased. PhaseNet has a similar tendency. We plan to study more optimal parameters and fine-tuning by adding training data around Japan.
Acknowledgments
We used PhaseNet, EQTransformer, and SeisBench libraries. We used the waveforms of JMA, NIED, universities, and institutions and the JMA unified earthquake catalog.
We used PhaseNet (Zhu and Beroza 2019) and EQTransformer (Mousavi et al. 2020). Both pickers were trained using global seismic waveforms to detect and classify P- and S-phases. We used the PF method (Tamaribuchi et al. 2016, 2018) for phase association and hypocenter determination. The PF method is currently applied to the automatic hypocenter determination of the JMA catalog.
To evaluate the performance in the normal and swarm period, we used the continuous waveforms at about 1300 stations nationwide for 12 hours each from 0:00 on March 1, 2011 and 12:00 on March 11, 2011. Since it takes a too long time to read the waveform data in WIN format directly with python, we converted it to mini SEED format and then performed phase pickers. As a result, it took about 2 hours and 30 minutes in the total workflow including the format conversion, picking, and hypocenter determination processing for the 12-hour waveforms.
As for the result of EQTransformer, we determined 278 events automatically, compared to 196 events in the JMA catalog for 12 hours from 0:00 on March 1, 2011. We also determined 862 events automatically, compared to 1192 events in the JMA catalog for 12 hours from 12:00 on March 11, 2011. Although the number of noises (false-positive) has decreased compared to the conventional PF method, the number of missing earthquakes (false-negative) has increased. PhaseNet has a similar tendency. We plan to study more optimal parameters and fine-tuning by adding training data around Japan.
Acknowledgments
We used PhaseNet, EQTransformer, and SeisBench libraries. We used the waveforms of JMA, NIED, universities, and institutions and the JMA unified earthquake catalog.