17:15 〜 18:45
[SSS04-P04] Earthquake Detection in the Taiwan MiDAS Borehole Seismometer Array
キーワード：microseismic, borehole seismometer, machine learning model
The 2018 Hualien earthquake caused severe damage in the Hualien region, Taiwan, along the Milun fault. Six years after the earthquake, the Milun Fault Drilling and All-inclusive Sensing project (MiDAS) initiated two scientific drillings in the hanging wall (700 m, Hole A) and the foot wall (500 m, Hole B) of the northern Milun fault. It successfully drilled through the Milun fault in Hole A and got a 50-m thick fault core. After drilling, the MiDAS project deployed an optical-fiber cable and borehole seismometer arrays in both holes to monitor seismicity on the Milun fault.
In this study, we use data recorded by the borehole seismometer arrays to establish an earthquake catalog, aiming to gain insight into the microseismicity on the Milun Fault. We apply two methods for seismic event picking; one is the regular manual picking method; another is a machine learning phase picking technique, RED-PAN (Liao et al., IEEE Trans. Geosci. Remote Sens., 2022). Our study period is from March 16 to April 15, 2023 (31 days). In the manually picking results, we identify 621 earthquakes. Among them, 45 microseismic events are strongly related to the Milun fault, which ts-tp time less than 1 second. On the other hand, the RED-PAN has been proven to be most efficient in processing the MiDAS data with a high-pass filter of 10Hz and a prediction time window of 30 seconds. The RED-PAN detects 830 events based on this configuration. 77% detections are verified as earthquakes. However, our results show that the RED-PAN is not familiar to detect microevents just near the sensors. Only 10 events with a ts-tp time less than 1 second are detected. The precision of 22% is relatively low compared to it of the detection for entire events. Our next step is to retrain the model of RED-PAN and increase its detection ability for nearby events. We expect that an updated RED-PAN is able to operate independently in event picking process and save workforce. It would be a huge benefit for further seismological studies in the MiDAS project.
In this study, we use data recorded by the borehole seismometer arrays to establish an earthquake catalog, aiming to gain insight into the microseismicity on the Milun Fault. We apply two methods for seismic event picking; one is the regular manual picking method; another is a machine learning phase picking technique, RED-PAN (Liao et al., IEEE Trans. Geosci. Remote Sens., 2022). Our study period is from March 16 to April 15, 2023 (31 days). In the manually picking results, we identify 621 earthquakes. Among them, 45 microseismic events are strongly related to the Milun fault, which ts-tp time less than 1 second. On the other hand, the RED-PAN has been proven to be most efficient in processing the MiDAS data with a high-pass filter of 10Hz and a prediction time window of 30 seconds. The RED-PAN detects 830 events based on this configuration. 77% detections are verified as earthquakes. However, our results show that the RED-PAN is not familiar to detect microevents just near the sensors. Only 10 events with a ts-tp time less than 1 second are detected. The precision of 22% is relatively low compared to it of the detection for entire events. Our next step is to retrain the model of RED-PAN and increase its detection ability for nearby events. We expect that an updated RED-PAN is able to operate independently in event picking process and save workforce. It would be a huge benefit for further seismological studies in the MiDAS project.