| 摘要: | 2018年0206花蓮地震因觸發了米崙斷層滑動，造成該斷層沿線的花蓮地區遭遇嚴重的災損。而該事件也引發國內外學界對其與米崙斷層構造的高度關注。此後的七年間已有多項相關研究陸續發表，顯示科學界對米崙斷層構造特性的重視。為進一步釐清米崙斷層的構造特性及花蓮地區地震相關性、潛在風險，中央研究院與中央大學地震風險管理中心（Earthquake Disaster & Risk Evaluation and Management Center，簡稱E-DREaM）共同執行了米崙斷層鑽井研究計畫（Milun Fault Drilling and All-inclusive Sensing project，簡稱MiDAS計畫）。該計畫分別於斷層上盤及下盤鑽取700公尺（井A）及500公尺（井B）的科學探測井，並成功於地表及井A、井B中佈署光纖地震儀與井下地震儀陣列，藉此監測米崙斷層的地震活動。本研究利用2023年3月16日至同年9月15日，共184天之MiDAS井下地震儀波形紀錄，建立地震目錄，以探討米崙斷層之微地震活動性。地震挑選方面，將採用人工辨識與深度學習模型RED PAN（Real-time Earthquake Detection and Phase-picking with multi-task Attention Network）兩種方法分別建立地震目錄，並根據S-P到時差及波傳方向對地震事件進行分類。
於此時段內，人工方法共識別出2,857筆地震事件，平均每日可偵測約16筆，其中21.8%為鄰近測站的地震（S-P到時差小於2秒）。進一步與氣象署地震目錄比對發現，鄰近測站的地震中約有95%之事件未記錄於氣象署地震目錄，顯示MiDAS井下地震儀對近場微地震具有優異的偵測能力。此外於鄰近測站的地震中，有306筆事件為鄰近測站的極淺層地震（震源深度至少小於500公尺，分類於S類）。本研究針對其中76筆發生於2023年5月2日之地震群進行深度分析，結果顯示該群地震之震源深度介於4至34公尺，推測為發生於排水溝下的自然極淺層地震，可能為液體流動所誘發的微地震事件。
而於深度學習模型－RED-PAN部分，本研究分別比較偵測時間視窗為60秒及30秒之模型的偵測表現。結果顯示，60秒視窗之模型（RED-PAN(60s)）具最佳表現，其精確率達76%、召回率78%。其中，對於區域型及遠場地震之召回率高達97%（S-P到時差大於2秒），但對鄰近測站之S類地震僅有3%，顯示RED-PAN模型對發生於感測器附近之微地震仍缺乏偵測能力。因此，解決RED-PAN模型對於鄰近事件之偵測能力為未來研究之首要課題，期望能發展出可長期穩定運行於MiDAS觀測系統的自動化處理系統，這將對於未來之二氧化碳封存及地熱計畫提供良好的技術支援。
本研究結果顯示花蓮北部地區地震活動度高，且未被地表地震網偵測到之微地震活動頻繁，未來搭配地震定位及規模估算，則可更完整地解析北花蓮地區微地震活動構造。而由人工地震偵測及深度學習地震偵測結果的差異顯示，若深度學習之模型沒有因區域而校正，則可能會忽略許多地震事件，這點值得給未來計畫使用深度學習模型進行地震偵測的研究借鏡。;The 2018 Hualien earthquake on Feb. 6th triggered coseismic slip along the Milun Fault, causing severe damage in Hualien City. To better understand the Milun Fault’s structure and associated seismic hazards, Institute of Earth Sciences, Academia Sinica and the Earthquake Disaster & Risk Evaluation and Management Center, National Central University initiated the Milun Fault Drilling and All-inclusive Sensing project (MiDAS). This project drilled two scientific boreholes: one was on the hanging wall (700 m, Hole A) and another was on the footwall (500 m, Hole B) of the northern Milun Fault. After drilling, an optical-fiber cable and borehole seismometer arrays were deployed in both boreholes and on the surface to monitor seismicity near the fault. In this study, we analyze seismic waveforms recorded by the borehole seismometers from March 16 to September 15, 2023 (184 days in total) to establish an earthquake catalog near the northern Hualien. We apply two methods to detect seismic events: the manual picking method and a deep learning phase picking technique, called RED PAN (Real-time Earthquake Detection and Phase-picking with multi-task Attention Network). Subsequently, we classify the seismic events based on the arrival time differences of S and P waves (Δts-p) and the directions of waveform propagation in the boreholes.
In the results of manual picking analysis, 2,857 seismic events have been detected during the study period. There are 16 events per day in average. Among these, 21.8% are near-station events (Δts-p < 2 s). About 95% of these near-station events are not detected by the Central Weather Administration (CWA). It indicates that the MiDAS borehole arrays have higher capability in earthquake detection compared to the regional surface network (CWASN). Furthermore, there are 306 events occurring at ultra-shallow depths (< 500 m) during the study period, called Class S. We obtain that a cluster of 76 Class S events on May 2, 2023 occurred along the storm drain nearby with focal depths ranging from 4 to 34 m. These events may be induced by fluid migration.
For the deep learning analysis, we compare RED-PAN models with two different time window lengths (60 seconds and 30 seconds). The model with a 60-second prediction window shows the best performance, achieving a precision of 76% and a recall of 78%. For regional events (Δts-p > 2 s), the recall reaches 97%, indicating the model′s strong ability to detect regional earthquakes. However, the recall for Class S is only 3%, suggesting that RED-PAN currently struggles to identify microevents occurring very close to the sensors. Enhancing the model′s capability to detect near-field events will be the priority in the future. The results would be useful for the Carbon Capture and Storage as well as geothermal projects.
In conclusion, our results indicate that the seismicity in the northern Hualien is high. Many microearthquakes have not been detected by the regional seismic network (CWASN). We may identify potential structures for these microevents by further estimating their locations and magnitudes in the future. Furthermore, the discrepancy of earthquake detections between the manual and deep learning methods suggests that the deep learning method would be missing amount of microevents if we do not retrain the model by local seismic data. This serves as a valuable reference for future research involving deep learning. |
