5:15 PM - 6:45 PM
[STT36-P05] Application of a deep-learning method to seafloor distributed acoustic sensing data off Sanriku for automatic picking of P- and S-waves
Keywords:Distributed acoustic sensing, Seafloor optical cable, Machine learning
Distributed acoustic sensing (DAS) is an emerging technology for seismic observation. DAS observation has great potential to improve the quality of event detection and hypocenter location due to its high spatial resolution. The DAS measurement utilizes a fiber-optic telecommunication cable as one-dimensional strain-meter array along the cable. DAS measurement using a seafloor cable enables seafloor observations with numerous sampling points. However, because DAS data has so many channels, it is difficult to pick up arrivals manually for all channels. PhaseNet-DAS (Zhu et al., 2023) has been developed as a deep-learning arrival time picker for DAS data. In this study, we use the PhaseNet-DAS to pick arrival times from seafloor DAS data using off-Sanriku cable observation system belonging to Earthquake Research Institute, The University of Tokyo.
The DAS data for this study were obtained using the QuantX DAS interrogator by OptaSense, part of Luna Innovations. The seafloor cable for the observation off Sanriku is approximately 100 km long. The observation period is from July 20 to July 31, 2023. The spatial sampling interval, the sampling frequency and the gauge length are ~5.1 m, 800 Hz, and ~51.0 m, respectively. During the observation period, 393 events were located by JMA in a region within a distance of ~111 km from the cable. The PhaseNet-DAS model used in this study was trained by Zhu et al. (2023) using onshore DAS data at Long Valley and Ridgecrest, California. In order to satisfy the requirements of this model, our offshore data were resampled at a spatial sampling interval of ~10.2 m and a sampling frequency of 100 Hz. Then we extracted records for 393 events. Each of the extracted records starts at an origin time from the JMA catalog and has a duration of 1 minute. The extracted data were read by PhaseNet-DAS to obtain arrival times of P- and S-waves. We applied no processing to the data after the extraction.
Arrival times of a large number of channels could be picked correctly for large events and events close to the cable. We confirmed that the PhaseNet-DAS read arrival times of P- and S-waves with an accuracy of less than 0.1 s for the best case by comparing automatic readings to manual readings, for the event shown in Fig. b. Additionally, there were almost no noises that were misinterpreted as P- or S- waves.
On the other hand, arrivals of P-waves were identified as those of S-waves, and vice versa for data with low signal-to-noise ratio. In addition, arrivals of converted waves with large amplitude were sometimes identified as those of P or S waves. Building a new PhaseNet-DAS model for offshore data may be effective to address these issues. However, it is challenging for now due to the lack of manually picked offshore data. In this study, PhaseNet-DAS performed well for data with high signal-to-noise ratio, so in order to improve signal-to-noise ratio, we tried high-pass filtering or channel-wise stacking for seismic records before automatic picking. However, neither of them was very effective.
Towards a precise hypocenter location using automatic reading by a deep learning method, it is important to evaluate the results and select correct arrival times from results. In addition, application of the automatic picking method to another oceanic region is also future work.
The DAS data for this study were obtained using the QuantX DAS interrogator by OptaSense, part of Luna Innovations. The seafloor cable for the observation off Sanriku is approximately 100 km long. The observation period is from July 20 to July 31, 2023. The spatial sampling interval, the sampling frequency and the gauge length are ~5.1 m, 800 Hz, and ~51.0 m, respectively. During the observation period, 393 events were located by JMA in a region within a distance of ~111 km from the cable. The PhaseNet-DAS model used in this study was trained by Zhu et al. (2023) using onshore DAS data at Long Valley and Ridgecrest, California. In order to satisfy the requirements of this model, our offshore data were resampled at a spatial sampling interval of ~10.2 m and a sampling frequency of 100 Hz. Then we extracted records for 393 events. Each of the extracted records starts at an origin time from the JMA catalog and has a duration of 1 minute. The extracted data were read by PhaseNet-DAS to obtain arrival times of P- and S-waves. We applied no processing to the data after the extraction.
Arrival times of a large number of channels could be picked correctly for large events and events close to the cable. We confirmed that the PhaseNet-DAS read arrival times of P- and S-waves with an accuracy of less than 0.1 s for the best case by comparing automatic readings to manual readings, for the event shown in Fig. b. Additionally, there were almost no noises that were misinterpreted as P- or S- waves.
On the other hand, arrivals of P-waves were identified as those of S-waves, and vice versa for data with low signal-to-noise ratio. In addition, arrivals of converted waves with large amplitude were sometimes identified as those of P or S waves. Building a new PhaseNet-DAS model for offshore data may be effective to address these issues. However, it is challenging for now due to the lack of manually picked offshore data. In this study, PhaseNet-DAS performed well for data with high signal-to-noise ratio, so in order to improve signal-to-noise ratio, we tried high-pass filtering or channel-wise stacking for seismic records before automatic picking. However, neither of them was very effective.
Towards a precise hypocenter location using automatic reading by a deep learning method, it is important to evaluate the results and select correct arrival times from results. In addition, application of the automatic picking method to another oceanic region is also future work.