5:15 PM - 7:15 PM
[SSS06-P01] Mutsu Bay Seismic Cluster along the Volcanic Front in November 2024
Keywords:Mutsu Bay, Phase picking, Hypocenter relocation, Machine learning
A sudden seismic activity began beneath Mutsu Bay in Aomori Prefecture, initiated by an Mj 4.6 earthquake at a depth of 8 km on 16 November 2024, according to the JMA unified catalog. Four days later, on 20 November, the largest earthquake, Mj 5.1, in this sequence occurred at a depth of 10 km. Its focal mechanism from the NIED F-net moment tensor solution showed WSW- or ENE-dipping nodal planes with a high CLVD component. Between 16 and 30 November, a total of 478 events with Mj 0.0 or greater were observed in this area. These events were clustered at 41.035° N and 141.070° E in a cylinder-like shape, extending 4 km laterally; however, their depth ranged from 6 to 14 km. Interestingly, this seismic cluster was situated along the volcanic front. Thus, it is important to understand what kind of faulting has occurred in this cluster. To tackle this question, we relocated catalog hypocenters to obtain a high-quality hypocenter distribution using machine learning techniques.
We selected 555 events with Mj 0.0 or greater from January 2004 to November 2024 for relocation, with 476 out of 478 events relocated after 9 August 2024. As the number of JMA picks were limited, we used a machine learning model to increase available phase picks. Thus, we conducted the following two steps: (1) picking of P- and S-wave arrivals for selected events using a deep-neural-network-based picker (Zhu and Beroza 2019) re-trained on the JMA unified catalog (Naoi et al. 2024) and (2) hypocenter relocation with picked onsets and waveform cross correlation (CC) values. We computed theoretical travel times and used them as reference. If the PhaseNet picker returns the probability 0.5 or greater, we used this onset time of the target phase; otherwise, we replaced it with the JMA pick if available. Using these picked onsets, we calculated waveform CC between adjacent event pairs. A seismic waveform in the window of 1s before and after the phase onset, filtered by 4–12 Hz, was used to calculate P- and S-wave correlation. Vertical and horizontal components were used for P-wave and S-wave correlation, respectively. For S wave, we computed the correlation of the horizontal particle motion by creating a complex waveform, instead of adopting the horizontal component with better CC values (e.g. Yoshida et al. 2023). The time shift that maximizes CC (0.8 or greater) was adopted. Creating differential travel times for picked onsets and CC, we applied hypoDD (Waldhauser and Ellsworth 2000) with the JMA2001 1-D velocity model (Ueno et al. 2002).
For PhaseNet-based phase picking, we obtained 18,709 picks for P wave and 14,458 picks for S wave, respectively, which doubled the number of JMA picks. We also obtained 25,6124 and 187,486 pairs for P- and S-waves as CC-based differential times, respectively.
The relocation results in 16–30 November showed that the lateral extension did not vary significantly, remaining less than 4 km wide; however, the depth extension was greatly limited to 9–10 km. The relocated hypocenters exhibited a disk-like shape rather than cylindrical one. Given the high CLVD component of large events, this cluster is not originated from active-fault earthquakes but could rather include fluid-induced or volcanic events. Further investigation into hypocenters, focal mechanisms, and seismic structures would advance our understanding of the local seismicity.
[Acknowledgments]
We used JMA Unified Catalog and phase picks. We analyzed seismic waveforms recorded by NIED Hi-net, JMA, ERI, Hokkaido, Tohoku, and Hirosaki Universities, and AS-net from the Association for the Development of Earthquake Prediction. This study was supported by MEXT Project for Seismology toward Research Innovation with Data of Earthquake (STAR-E) Grant Number JPJ010217.
We selected 555 events with Mj 0.0 or greater from January 2004 to November 2024 for relocation, with 476 out of 478 events relocated after 9 August 2024. As the number of JMA picks were limited, we used a machine learning model to increase available phase picks. Thus, we conducted the following two steps: (1) picking of P- and S-wave arrivals for selected events using a deep-neural-network-based picker (Zhu and Beroza 2019) re-trained on the JMA unified catalog (Naoi et al. 2024) and (2) hypocenter relocation with picked onsets and waveform cross correlation (CC) values. We computed theoretical travel times and used them as reference. If the PhaseNet picker returns the probability 0.5 or greater, we used this onset time of the target phase; otherwise, we replaced it with the JMA pick if available. Using these picked onsets, we calculated waveform CC between adjacent event pairs. A seismic waveform in the window of 1s before and after the phase onset, filtered by 4–12 Hz, was used to calculate P- and S-wave correlation. Vertical and horizontal components were used for P-wave and S-wave correlation, respectively. For S wave, we computed the correlation of the horizontal particle motion by creating a complex waveform, instead of adopting the horizontal component with better CC values (e.g. Yoshida et al. 2023). The time shift that maximizes CC (0.8 or greater) was adopted. Creating differential travel times for picked onsets and CC, we applied hypoDD (Waldhauser and Ellsworth 2000) with the JMA2001 1-D velocity model (Ueno et al. 2002).
For PhaseNet-based phase picking, we obtained 18,709 picks for P wave and 14,458 picks for S wave, respectively, which doubled the number of JMA picks. We also obtained 25,6124 and 187,486 pairs for P- and S-waves as CC-based differential times, respectively.
The relocation results in 16–30 November showed that the lateral extension did not vary significantly, remaining less than 4 km wide; however, the depth extension was greatly limited to 9–10 km. The relocated hypocenters exhibited a disk-like shape rather than cylindrical one. Given the high CLVD component of large events, this cluster is not originated from active-fault earthquakes but could rather include fluid-induced or volcanic events. Further investigation into hypocenters, focal mechanisms, and seismic structures would advance our understanding of the local seismicity.
[Acknowledgments]
We used JMA Unified Catalog and phase picks. We analyzed seismic waveforms recorded by NIED Hi-net, JMA, ERI, Hokkaido, Tohoku, and Hirosaki Universities, and AS-net from the Association for the Development of Earthquake Prediction. This study was supported by MEXT Project for Seismology toward Research Innovation with Data of Earthquake (STAR-E) Grant Number JPJ010217.