利用環境資料預測地表位移及坡地災害事件

Abstract

自1940年代起，隨著全球氣溫逐年升高，災害發生頻率也顯著上升。重大災害事件數從每年約20起增加至每年超過400起，顯示全球均溫上升驅動了更多致災因子，直接反映在災害事件數量上及次生災害之比率。劇烈天氣事件能改變近地表地質特性、觸發地質災害事件如地層下陷、土壤液化、坡地災害等，由於這些地質災害事件可反應在地表變形資料上，如何利用環境資料（包含大氣類、地下水、潮汐等資料）預測地表變形資料（GNSS和地震資料）及坡地災害事件？這幾個問題驅動了我們對「天然災害鏈預測」可行性評估的動機。本研究的主要工作目標為利用長短期記憶模型（LSTM模型）及支持向量迴歸（SVR模型）之演算法，建立地下水位、地表變形及坡地災害之預測模型。首先我們利用2004年至2020年潮州地區兩個自動氣象站所提供的氣溫、雨量、風向、風速，以及目標預測測站之歷史地下水位，預測十二個地下水測站的地下水位，比較LSTM模型和SVR模型的表現，結果顯示，大部分地下水測站的LSTM模型在測試集上的平均決定係數高達0.90。其次，我們應用相同方法，探討2013年至2020年垂直地表變形預測，新加入了氣壓、相對濕度以及潮位資料，發現GNSS測站的LSTM模型在測試集上的平均決定係數達0.94，而更高採樣點率寬頻地震站位移場的平均決定係數達0.89，不同測試皆證實LSTM較SVR模型預測表現更佳。最後，我們以農業部農村發展及水土保持署的坡地災害事件目錄為基礎，評估了不同年份組成的四個資料集。由2010年至2012年的訓練集以及2013年至2014年的測試集組成的資料集表現最佳，顯示了環境資料在坡地災害預測中的潛力，準確率達0.83，精準率為0.95，召回率為0.67。綜合以上， LSTM模型展示了對於時序資料強大的預測能力，並強調了環境資料在地表變形及坡地災害預測中的關鍵作用。本研究並測試六個不同的氣象參數在預測模型之貢獻度，依照重要程度依序如下：氣溫、氣壓、風速、風向、相對濕度、雨量，然而同時考慮此六個氣象參數的模型，其預測效能仍優於單一氣象參數。本研究具體提供了氣象與地質災害間的預測方法論，期能在未來用於近即時警報/預報、並為未來政策制定提供即時參考依據。Since the 1940s, global temperatures have been gradually rising, correlating with a significant increase in the frequency of nature disasters. The number of major disasters has increased from ~20 per year to over 400 annually, indicating the impact of global warming. Extreme weather events have been discovered to significantly alter near-surface geological properties, leading to geological hazards including land subsidence, soil liquefaction, landslide, slopeland, and so on. The main objective of this study is to develop predictive models of groundwater levels, surface deformation, and slopeland disasters in southwestern Taiwan near Chaozhou. The study area is chosen due to the high subsidence rate and low seismicity. We attempt to establish the possibility of predicting surface deformation data (GNSS and seismic data) and geological hazards using environmental data (i.e., atmospheric, groundwater, and tidal data) and machine learning approaches (Long Short-Term Memory, LSTM and Support Vector Regression, SVR). In the study period of 2004 to 2020, this study initially utilized temperature, rainfall, wind direction, and wind speed data from two automatic weather stations in the study area to predict groundwater levels at twelve various groundwater stations. The resulting prediction performance (averaged coefficient of determination) reached 0.90 at most of groundwater stations using LSTM model. We next applied the same method to explore possibility of vertical surface deformation prediction incorporating additional data such as air pressure, relative humidity, and tide levels. We found that at the targeted GNSS stations, the averaged coefficient of determination up to 0.94. At the broadband seismometer displacement field characterized by much finer time resolution, the coefficient of determination reached 0.89. Various tests confirmed that the LSTM model outperformed the SVR model in prediction accuracy. Finally, we used four sets of environmental data from a variety of data period to predict slopeland disasters. We found that as long as the particular years experienced extreme landslide events were excluded in the training data, the high prediction performance can be reached. The best model reveals the accuracy of 0.83, precision of 0.95, and recall of 0.67. In conclusion, LSTM models showed robust predictive capabilities for time-series data that highlights the pivotal role of environmental data in forecasting surface deformation and landslide events. In the future, developing predictive models for various geological hazard types can be expected with the hope of offering timely warnings and predictions for natural disaster prevention.