5:15 PM - 6:45 PM
[SCG50-P09] Application of seismic detection techniques based on deep learning: Toward practical use of inland low-frequency earthquakes
Keywords:Deep-Learning, Low-frequency earthquakes, Phase picking
Low-frequency earthquakes (LFEs) are those in which low frequencies dominate despite the small magnitude. Due to the characteristics of their waveforms and low S/N, determining the arrival time of seismic waves radiated by LFEs is challenging rather than regular earthquakes. On the other hand, in recent years, with the development of deep learning, there have been significant advances in automatic seismic wave detection techniques. However, there are still not many examples of the application of this technique in Japan, especially for LFEs that occur inland (inland LFEs), mostly nearby volcanoes. In this study, we attempted to apply the deep-learning-based technique to both regular earthquakes and inland LFEs.
In this study, PhaseNet developed by Zhu and Beroza (2019) is used as a seismic wave arrival detection model. The model takes three-component seismic waveforms as input and outputs the arrival times of P- and S-waves as probabilities with respect to time. We analyzed earthquakes whose magnitude of 2-4 occurred inland with depths shallower than 30 km and LFEs, which have occurred in the northern Tohoku region from April 2004 to December 2023, based on the JMA unified earthquake catalog. Notice that LFEs were selected based on the flags in the JMA catalog. We used velocity waveform records at Hi-net stations of the National Research Institute for Earth Science and Disaster Prevention. The 40 seconds of three-component records starting from 10 seconds before the origin time were used as input to the PhaseNet. For inland earthquakes, raw observed traces were used, while for LFEs, five types of filters with the following passbands were applied in addition to the raw traces: bandpass filters of 0.5-1, 1-2, and 2-4 Hz, and high-pass filters of 1 Hz and 2 Hz. A total of six types of three-component waveforms were input into the model for the case of LFEs to pursue better detections of the first arrival. For both of earthquake types, the time of the peak value of the probability was regarded as the arrival times. In addition for the case of a LFEs, if the peak probability values of both P and S waves in the raw waveforms were less than 0.5, the peak value with the highest probability from the six filtered waveforms was selected. Then, we compared our travel times with cataloged values at the same stations.
We confirmed that the measured travel times for regular earthquakes were well consistent with those in JMA's catalog. On the other hand for LFEs, detection of the raw waveforms was somewhat difficult but there were cases when could be improved by filter processing. The high-pass filter with the corner frequency of 2 Hz was particularly effective. The peak probability values for LFEs were rarely as high as those for regular earthquakes. Besides, we found that PhaseNet tended to read slightly backward in time as arrival times compared to the JMA measure values for both regular earthquakes and LFEs. This may be a bias caused by using data from a specific region for PhaseNet training. In the future, it is expected that re-training with seismic waveforms observed in Japan and applying noise reduction processing by machine learning to records with low S/N ratio will enable us to obtain highly accurate source distribution of LFEs.
In this study, PhaseNet developed by Zhu and Beroza (2019) is used as a seismic wave arrival detection model. The model takes three-component seismic waveforms as input and outputs the arrival times of P- and S-waves as probabilities with respect to time. We analyzed earthquakes whose magnitude of 2-4 occurred inland with depths shallower than 30 km and LFEs, which have occurred in the northern Tohoku region from April 2004 to December 2023, based on the JMA unified earthquake catalog. Notice that LFEs were selected based on the flags in the JMA catalog. We used velocity waveform records at Hi-net stations of the National Research Institute for Earth Science and Disaster Prevention. The 40 seconds of three-component records starting from 10 seconds before the origin time were used as input to the PhaseNet. For inland earthquakes, raw observed traces were used, while for LFEs, five types of filters with the following passbands were applied in addition to the raw traces: bandpass filters of 0.5-1, 1-2, and 2-4 Hz, and high-pass filters of 1 Hz and 2 Hz. A total of six types of three-component waveforms were input into the model for the case of LFEs to pursue better detections of the first arrival. For both of earthquake types, the time of the peak value of the probability was regarded as the arrival times. In addition for the case of a LFEs, if the peak probability values of both P and S waves in the raw waveforms were less than 0.5, the peak value with the highest probability from the six filtered waveforms was selected. Then, we compared our travel times with cataloged values at the same stations.
We confirmed that the measured travel times for regular earthquakes were well consistent with those in JMA's catalog. On the other hand for LFEs, detection of the raw waveforms was somewhat difficult but there were cases when could be improved by filter processing. The high-pass filter with the corner frequency of 2 Hz was particularly effective. The peak probability values for LFEs were rarely as high as those for regular earthquakes. Besides, we found that PhaseNet tended to read slightly backward in time as arrival times compared to the JMA measure values for both regular earthquakes and LFEs. This may be a bias caused by using data from a specific region for PhaseNet training. In the future, it is expected that re-training with seismic waveforms observed in Japan and applying noise reduction processing by machine learning to records with low S/N ratio will enable us to obtain highly accurate source distribution of LFEs.