9:30 AM - 9:45 AM
[SCG52-03] Deep-neural-network-based arrival-time picking and earthquake location: application to S-net
Keywords:S-net, Neural networks, Computational seismology, Seismic monitoring
1. Introduction
Off the Pacific coast of Tohoku Japan, the new cabled ocean bottom seismic network (S-net) gives us a large number of seismograms of offshore earthquakes. Automatic arrival time picking (hereafter, phase picking) and event location methods make an important contribution for fast earthquake detection and location from such a large dataset. In our previous study, we evaluated a phase picking method based on a deep neural network, PhaseNet (Zhu and Beroza, 2019) for application to S-net (Suzuki et al., 2020: JpGU-AGU Joint Meeting). Although we used the pre-trained model with the seismic data in northern California for the phase picking, PhaseNet showed comparable or better performance compared with two other automatic phase picking methods. In this study, we train PhaseNet using S-net data and estimate arrival times with the newly trained PhaseNet. We then, associate the phases to detect events and estimate their hypocenters.
2. Data and method
The S-net has been deployed since 2013 by the National Research Institute for Earth Science and Disaster Resilience, and the continuous records from 2016 are available. We focused on a M5.6 event on Aug. 8. 2018 (hereafter, the mainshock) that occurred near the Japan trench offshore Miyagi. Note that the S-net data has been used for JMA’s hypocenter location since September 2020. We first trained the PhaseNet with data from 150 S-net and 492 other land stations by 727 events. We used 98,766 manual picks from onshore (39,858) and offshore (58,918) stations in the Tohoku region. The number of the P and S wave picks are the same. After training PhaseNet, we used continuous waveform data from 9 S-net stations near the mainshock for phase picking by the newly trained PhaseNet for the period from 1 week before to 2 weeks after the mainshock. We then performed earthquake detection by associating the picked phases to individual events using a phase association method (REAL; Zhang et al., 2019) to find new events offshore Tohoku. This method REAL associates P and S phases from multiple stations with a single event and locates its hypocenter with a grid search to minimize the travel time residuals. We searched potential hypocenters shallower than 50 km and within the area of 0.48 × 0.48 degrees area centered at the station that recorded the first-arriving phase for each events. Finally, we refined the hypocenter location by a nonlinear earthquake location method (hypomh; Hirata and Matsu’ura, 1987).
3. Results
The newly trained PhaseNet using S-net and land stations data achieved good performance (F1 score: harmonic mean of precision and recall rate) for both P (0.745) and S (0.494) waves. This F1 score is better than that of the original model (0.409 for P waves and 0.283 for S waves) that was trained by the northern California data. This suggests despite the smaller number (1/10) of training data, the new model trained with S-net data has better performance for S-net data.
Using the newly trained model of PhaseNet and the phase association method (REAL), we detected 106 events in the region during the three-week period while JMA listed 43 events using land data only. This suggests the sequential application of PhaseNet and REAL is potentially effective to automatically detect new earthquakes. We expect that the phase picking method based on a deep neural network will contribute to revealing the detailed seismicity offshore Tohoku, Japan by using S-net waveform data.
Off the Pacific coast of Tohoku Japan, the new cabled ocean bottom seismic network (S-net) gives us a large number of seismograms of offshore earthquakes. Automatic arrival time picking (hereafter, phase picking) and event location methods make an important contribution for fast earthquake detection and location from such a large dataset. In our previous study, we evaluated a phase picking method based on a deep neural network, PhaseNet (Zhu and Beroza, 2019) for application to S-net (Suzuki et al., 2020: JpGU-AGU Joint Meeting). Although we used the pre-trained model with the seismic data in northern California for the phase picking, PhaseNet showed comparable or better performance compared with two other automatic phase picking methods. In this study, we train PhaseNet using S-net data and estimate arrival times with the newly trained PhaseNet. We then, associate the phases to detect events and estimate their hypocenters.
2. Data and method
The S-net has been deployed since 2013 by the National Research Institute for Earth Science and Disaster Resilience, and the continuous records from 2016 are available. We focused on a M5.6 event on Aug. 8. 2018 (hereafter, the mainshock) that occurred near the Japan trench offshore Miyagi. Note that the S-net data has been used for JMA’s hypocenter location since September 2020. We first trained the PhaseNet with data from 150 S-net and 492 other land stations by 727 events. We used 98,766 manual picks from onshore (39,858) and offshore (58,918) stations in the Tohoku region. The number of the P and S wave picks are the same. After training PhaseNet, we used continuous waveform data from 9 S-net stations near the mainshock for phase picking by the newly trained PhaseNet for the period from 1 week before to 2 weeks after the mainshock. We then performed earthquake detection by associating the picked phases to individual events using a phase association method (REAL; Zhang et al., 2019) to find new events offshore Tohoku. This method REAL associates P and S phases from multiple stations with a single event and locates its hypocenter with a grid search to minimize the travel time residuals. We searched potential hypocenters shallower than 50 km and within the area of 0.48 × 0.48 degrees area centered at the station that recorded the first-arriving phase for each events. Finally, we refined the hypocenter location by a nonlinear earthquake location method (hypomh; Hirata and Matsu’ura, 1987).
3. Results
The newly trained PhaseNet using S-net and land stations data achieved good performance (F1 score: harmonic mean of precision and recall rate) for both P (0.745) and S (0.494) waves. This F1 score is better than that of the original model (0.409 for P waves and 0.283 for S waves) that was trained by the northern California data. This suggests despite the smaller number (1/10) of training data, the new model trained with S-net data has better performance for S-net data.
Using the newly trained model of PhaseNet and the phase association method (REAL), we detected 106 events in the region during the three-week period while JMA listed 43 events using land data only. This suggests the sequential application of PhaseNet and REAL is potentially effective to automatically detect new earthquakes. We expect that the phase picking method based on a deep neural network will contribute to revealing the detailed seismicity offshore Tohoku, Japan by using S-net waveform data.