南海北部海床逸氣的氣泡訊號：應用機器學習分析海底地震儀紀錄

Abstract

海床逸氣是一種常見的地質現象，氣體來源可能來自地層中天然氣水合物的解離以及火成活動的氣體釋放，主要組成為烷類和二氧化碳，並由其來源處沿著地層裂隙移棲至海水層，最終逸散到大氣中，這個過程除了可能加劇全球氣候變遷之外，也會影響海床的穩定性，進而增加海域地質災害的風險。目前海床逸氣的紀錄主要來自研究船上其他觀測的副產物，例如多音束測深儀在探勘地形時所觀察到的海床氣焰，或是使用水下探測載具時直接觀察到海床逸散的氣泡。然而，如何有效監測海床逸氣在空間和時間上的變化，並探討其控制因子及評估相關風險，實為亟待解決的挑戰。本研究旨在透過海底地震儀的紀錄，分析海床逸氣所產生的氣泡訊號。氣體沿著地層裂隙移棲至沉積物與海水層的交界處，其膨脹所產生的微小爆裂會以地震SH波的形式傳遞到附近的海底地震儀，產生高頻率、單頻振動、衰減快且時長約1到4秒的訊號。本研究於2007、2008和2011年佈放在南海北部的海底地震儀紀錄中發現大量的氣泡訊號，故使用以長短期記憶神經網路作為主要架構的機器學習模型，從原始時序資料提取原始波形、時頻圖和波形包絡線等特徵進行訓練並辨識氣泡訊號以進行量化分析。結果顯示，大陸斜坡區域的氣泡數量與水深呈負相關，推測其原因為垂直荷重的差異。但在水深超過3500公尺的海盆中，多數氣泡訊號集中在海盆東側，可能與板塊隱沒的彎曲所造成之裂隙和此區域的火成活動有關。本研究也發現沿著測站佈設方向其訊號峰值有逐漸延遲的趨勢，原因可能為一具方向性的擾動，推測與海洋內部波動如內波有關。此外，氣泡訊號的峰值出現時間於部分測站與潮汐週期相關，顯示潮汐所造成的壓力差也可能觸發海床逸氣。綜合以上，本研究成功透過機器學習快速且自動化辨識海底地震儀中的海床逸氣訊號，並藉由觀察其在時空上之變化探討其控制因子，使大範圍的監測成為可能，未來將可有助於氣體排放監測以及區域地質災害的風險評估。

Seafloor gas emissions, mainly composed of alkanes and carbon dioxide, are a common geological phenomenon. These gases, typically originating from the dissociation of gas hydrates or volcanic activities, may migrate through fractures into seawater before eventually escaping into the atmosphere. The released gases and the entire process may exert a significant influence on global climate and impact on seafloor stability, increasing the risk of geological disasters. The observation of seafloor gas emissions is typically derived from shipboard surveys, such as using multibeam echosounders to observe gas flares or employing underwater remotely operated vehicles to directly observe gas bubbles. However, such approach may become impractical when continuous and comprehensive monitoring is required to observe the spatial and temporal variations of seafloor gas emissions. This study aims to analyze the bubble signals generated by seafloor gas emissions using records from ocean bottom seismometers (OBS). When gas migrates through fractures into the interface between sediment and seawater, the expansion results in tiny bursts that propagate as seismic SH waves to nearby OBS stations, producing high and single-frequency oscillations with rapid decay and a duration of approximately 1 to 4 seconds. Numerous bubble signals were discovered in the OBS records deployed in the northern South China Sea during 2007, 2008, and 2011. To address this, a machine learning model based on Long Short-Term Memory (LSTM) algorithm was utilized by using the extracted features from the original time-series data, allowing for the identification of bubble signals and the follow-up discussion. The results show that the amount of bubble signals on the continental slope decreased with the increase of water depth, likely due to variations in vertical loading. However, in the deeper abyssal basins exceeding 3500 meters, most bubble signals were concentrated on the eastern side, possibly linked to tension fractures caused by subducting plate bending or volcanic activities in the region. The research also revealed that a gradual time delay in peak occurrence along the direction of OBS deployment was observed, possibly attributed to directional disturbances associated with internal oceanic waves, such as internal waves. Moreover, the peak occurrence time of bubble signals at certain sites correlated with tidal cycles, suggesting that tidal-induced pressure differences could trigger seafloor gas emissions. In conclusion, this study effectively utilized machine learning to automatically identify seafloor gas emission signals from OBS data, shedding new light on their spatiotemporal distributions and controlling factors in the northern South China Sea. These findings provide valuable insights for monitoring seafloor gas emissions and understanding their environmental implications, which are crucial for the related risk assessments.