12:00 PM - 12:15 PM
[SCG55-06] Comprehensive detection of S wave later phase with deep learning: Application for a seismic swarm around the Moriyoshi volcano
Keywords:Seismic wave, S wave later phase, Machine learning, Seismic swarm
The development of automatic seismic wave processing techniques using machine learning has been accelerated, enabling accurate and exhaustive earthquake detection and phase picking (e.g., Zhu & Beroza, 2019). However, application to seismic later phases, such as reflected waves caused by significant velocity differences in the subsurface, has not progressed far enough. (Ding et al., 2022). Information on the crustal heterogeneity can be extracted from later phases. Therefore, automation of later phase detection or arrival time picking will enable us detailed and comprehensive analysis of the crustal heterogeneity.
We have developed a deep learning convolutional neural network model for automatic detection of later phase in S coda (LPS) (Amezawa et al., 2022, SSJ) using seismograms of an earthquake swarm around the Moriyoshi volcano in the northern Tohoku, Japan where clear LPS in S coda have been reported. In this area, a swarm activity has continued in the same cluster for more than 10 years since March 2011. In addition, a Hi-net station in the vicinity of the swarm area enables us to conduct continuous analysis of the characteristics of LPS. The comprehensive analysis of LPS over a long term is important to elucidate the nature of LSP origin. In this study, we performed comprehensive detection of LPS in this area using our detection model and discussed the result with the hypocenter distribution.
Seismograms of 9,779 earthquakes (M>=0) recorded from March 2011 to September 2022 at a Hi-net station (N.ANIH), located about 10 km southwest of the cluster, and S wave arrival times by JMA were used. We have used the waveform data up to January 2015 (5,000 seismograms) for model training and test. Thus, in this study, we automatically detected LPS using earthquakes since February 2015 and our trained model. The seismograms input to the model were transverse components and filtered by 6-24 Hz band-pass filter, and the length of the data is 5.12 s (0.2 s before the S wave arrival). We also relocated the hypocenters distribution by the double difference algorithm (Waldhauser & Ellsworth, 2000) for the discussion.
We set the threshold for LPS detection as the model output value above 0.5 and LPS was newly detected for 3,666 earthquakes. Together with our manual detection for model training, LPS was detected for 7,601 earthquakes. We found that undetected results were particularly common on the northeast side of the swarm cluster (Figure 1(a)). In addition, LSP was detected stably throughout the entire period, with no temporal characteristics such as the absence of detection for a certain period (Figure 1(b)).
Since no clear relationship was found between the magnitude of the earthquakes and the detection results, whether LPS is observable or not is considered to depend on the hypocenter location. This suggests a specific geometry in which LPS is difficult to generate. However, the LSP detections were not particularly localized. The reasons for the coexistence of detected and undetected results in the northeastern part may be spatiotemporal changes in the reflection characteristics of the LPS origin (due to fluid movement, etc.) or slight differences in the S wave radiation pattern caused by the differences in focal mechanisms.
Acknowledgements: We used seismic waveform records from NIED Hi-net, JMA unified hypocenter catalog, and arrival time. This research was supported by the Seismology TowArd Research innovation with data of Earthquake (STAR-E) Project [JPJ010217] of MEXT
We have developed a deep learning convolutional neural network model for automatic detection of later phase in S coda (LPS) (Amezawa et al., 2022, SSJ) using seismograms of an earthquake swarm around the Moriyoshi volcano in the northern Tohoku, Japan where clear LPS in S coda have been reported. In this area, a swarm activity has continued in the same cluster for more than 10 years since March 2011. In addition, a Hi-net station in the vicinity of the swarm area enables us to conduct continuous analysis of the characteristics of LPS. The comprehensive analysis of LPS over a long term is important to elucidate the nature of LSP origin. In this study, we performed comprehensive detection of LPS in this area using our detection model and discussed the result with the hypocenter distribution.
Seismograms of 9,779 earthquakes (M>=0) recorded from March 2011 to September 2022 at a Hi-net station (N.ANIH), located about 10 km southwest of the cluster, and S wave arrival times by JMA were used. We have used the waveform data up to January 2015 (5,000 seismograms) for model training and test. Thus, in this study, we automatically detected LPS using earthquakes since February 2015 and our trained model. The seismograms input to the model were transverse components and filtered by 6-24 Hz band-pass filter, and the length of the data is 5.12 s (0.2 s before the S wave arrival). We also relocated the hypocenters distribution by the double difference algorithm (Waldhauser & Ellsworth, 2000) for the discussion.
We set the threshold for LPS detection as the model output value above 0.5 and LPS was newly detected for 3,666 earthquakes. Together with our manual detection for model training, LPS was detected for 7,601 earthquakes. We found that undetected results were particularly common on the northeast side of the swarm cluster (Figure 1(a)). In addition, LSP was detected stably throughout the entire period, with no temporal characteristics such as the absence of detection for a certain period (Figure 1(b)).
Since no clear relationship was found between the magnitude of the earthquakes and the detection results, whether LPS is observable or not is considered to depend on the hypocenter location. This suggests a specific geometry in which LPS is difficult to generate. However, the LSP detections were not particularly localized. The reasons for the coexistence of detected and undetected results in the northeastern part may be spatiotemporal changes in the reflection characteristics of the LPS origin (due to fluid movement, etc.) or slight differences in the S wave radiation pattern caused by the differences in focal mechanisms.
Acknowledgements: We used seismic waveform records from NIED Hi-net, JMA unified hypocenter catalog, and arrival time. This research was supported by the Seismology TowArd Research innovation with data of Earthquake (STAR-E) Project [JPJ010217] of MEXT