深度學習地震監測系統SeisBlue應用在三個不同地震網的地震目錄

Abstract

近年來，地震監測的應用越來越廣泛，不僅包含對地震災害的研究，像是地震序列來研究地震周期、孕震構造，或者是最近的重要議題：淨零排碳中的碳封存和地熱能源的探勘與開發，可以對碳封存時注入地下的二氧化碳與地熱電廠注入地下的水引起的地震起到提前警示的作用，也可以對二氧化碳的移棲情形或地下熱液的位置做觀測。地震監測的除了有很多的應用外，隨著時間進展，在世界各地布放的地震儀越來越多，儀器也越來越進步，所記錄到的地震資料的累積速度也越來越快，以人工的方式來處理地震會面臨效率嚴重不足的問題。因此對於地震資料處理的各項步驟已有很多研究研發不同的自動化的軟體來節省時間和人力。使用深度學習模組的地震監測系統SeisBlue，是由地震資料經過深度學習模型挑波、關聯、定位、建立地震目錄的自動化系統，而其中的深度學習模型是利用台灣的地震資料經過人工挑選的資料訓練的。本研究使用SeisBlue對三種不同地震觀測網的資料建立地震目錄，分別是在北台灣的Formosa Array、在池上地區的池上寬頻地震網和臨時密集地震網，用以了解SeisBlue對不同的測站數量、測站密度、測站分布、測站儀器的地震觀測網的處理效能。在實際處理資料時，根據測站網與地震分布調整所使用的參數是很重要的。SeisBlue在Formosa Array 2020年的資料，共偵測出5955筆地震；在池上寬頻地震網 2022/9/1~2022/10/31的資料共偵測出14462筆地震；在池上臨時密集地震網 2022/9/18~2022/10/25共偵測出45224筆地震。其中Formosa Array網內在2020期間沒有發生較大的地震，而池上地震網與臨時密集地震網有包含2022/9/18芮氏規模 6.9 池上地震的餘震序列。SeisBlue與中央氣象署的地震目錄一起做比較，SeisBlue和中央氣象署定位的地震位置分布相近但是能偵測到更多的數量，使用臨時密集地震網可以比池上寬頻地震網偵測出更小的地震，對於背景地震序列和餘震序列，SeisBlue都可以有效率的偵測地震，另外使用更加密集的地震網則可以偵測到更小的地震。本研究也成功將SeisBlue應用於近即時資料處理系統，可以發布近即時地震報告，未來將可以應用在其他地區方便做即時地震監測，例如對碳封存區域的誘發地震監測。

Recently, the applications of earthquake monitoring are growing, which include the research of earthquake hazard, for example using earthquake sequences to find out earthquake cycle and the seismic structure. Earthquake monitoring is also necessary to the recently important issues, carbon capture and storage (CCS) and geothermal exploration. For CCS process, earthquake monitoring helps the observation of induced earthquakes and CO¬¬¬2 migration; for geothermal exploration, it helps finding out the geothermal reservoir and also the observation of induced earthquakes while geothermal power plant inject water. With more applications of earthquake monitoring, there are more and more seismometers placed around the world with better equipment, the accumulated seismic data grows faster and faster. Manual process can’t manage such great amount of data in efficiency. So, there are many studies have developed different automatic methods to decrease time and labor cost. SeisBlue, an earthquake monitoring system with Taiwan-data-training deep-learning model, can get earthquake catalog from earthquake data automatically, through the process of picking of deep-learning model, association, and locating. This research applies SeisBlue to three different seismic network to extract earthquake catalog, which is Formosa Array in northern Taiwan, Chihshang Seismic Network around Chihshang, and temporary dense nodal array around Chihshang, for the purpose of understanding the result of using SeisBlue to extract seismic catalog from different seismic network, with different station number, density, distribution, and instrument. When processing the data, adjusting the parameters according to the station and earthquake distribution is very important. As the result, SeisBlue extract 5955 events from Formosa Array within 2020; 14462 events from Chihshang Seismic Network during 2022/9/1~2022/10/31, 45224 events from Chihshang dense nodal array during 2022/9/18~2022/10/25. There is not large earthquake occurred during 2020 inside the Formosa Array, in contrast, Chihshang seismic network and dense nodal array contain the 2022/9/18 ML 6.9 Chihshang earthquake sequence. Compare to the catalog announced by Central Weather Administrator, the earthquake distribution of SeisBlue catalog has similar pattern to Central Weather Administrator catalog, but detect more earthquakes. With denser station distribution, SeisBlue can detect smaller earthquake and can detect earthquake efficiently both on background and aftershock seismicity. We also successfully apply SeisBlue to near realtime process system and can announce near realtime earthquake report, it can be used for realtime earthquake monitoring in the future, for example, the induced earthquake observation of the CCS region.