以機器學習建立邊坡滑動敏感性與降雨閥值模型用於崩塌災害早期預警系統

Abstract

由於位處板塊交界處與副熱帶季風氣候區，造就臺灣具有複雜多變的地質條件與變化劇烈的氣候。基於這樣的自然環境條件之下，導致山坡地區成為崩塌與地滑災害頻發的高風險區域，因此從中可得知完善的災前預警機制，對於臺灣山區具有極為重要的意義。因此本研究旨在建立一套機器學習模型，用於評估山坡崩塌潛勢，並且依據歷年崩塌災害事件的降雨資訊，創建區域內降雨門檻值曲線，用於降雨崩塌災害早期預警。 本研究以高屏溪集水區作為研究區域，藉由統整歷年崩塌目錄，劃設崩塌地區周圍之斜坡單元，並蒐集各單元共17項崩塌潛勢因子，作為建立機器學習模型所需的訓練與測試資料。研究中分別採用隨機森林、極限梯度提升以及集成學習三種機器學習模型，建構崩塌潛勢評估模型，並以混淆矩陣針對測試資料集進行模型準確度驗證。最終以預測表現最佳者，針對歷年災害事件進行潛勢分析，進而劃設50%、75%與90%崩塌潛勢下的臨界降雨門檻值曲線，建立具風險分級特性且可應用於實際災害預警的降雨門檻值判定模式。除此之外，本研究也以莫拉克與凱米颱風案例驗證崩塌潛勢模型，並且以發生於莫拉克颱風以及盧碧颱風的崩塌地滑案例佐證降雨門檻值曲線於實際應用上之可行性。 根據研究結果顯示，三種機器學習模型在測試資料的混淆矩陣驗證中，皆可達到高於85%的整體準確率，其中以集成學習模型表現最佳，並且三者皆採取保守評估策略，使其在真實應用中降低漏報情況發生的風險。另一方面，在莫拉克颱風案例的降雨預警分析中，本研究所擬定的具風險分級之降雨門檻值曲線，可於災害事件發生前的十小時內發出預警訊號。至於在盧碧颱風事件中，則觀察到門檻值曲線易受到降雨型態的影響，因此可令平均降雨強度與累積降雨之臨界門檻值互為雙重驗證，以此降低誤報風險。

Taiwan's position at the intersection of tectonic plates and within a subtropical monsoon climate zone results in highly complex geological conditions and frequent extreme rainfall events. These environmental factors render mountainous regions particularly vulnerable to landslides. To reduce disaster risk, establishing an effective pre-disaster warning system is essential. This study aims to develop a machine learning model for assessing landslide susceptibility and to construct regional rainfall threshold curves for early warning of rainfall-induced landslides. The Kaoping River watershed was selected as the study area. A historical landslide inventory was compiled, and slope units surrounding the landslide sites were delineated. For each unit, 17 landslide influential factors were collected to form the dataset for model training and testing. Three machine learning algorithms—Random Forest, Extreme Gradient Boosting, and ensemble learning—were applied to develop susceptibility models. Model performance was evaluated using confusion matrix analysis on the testing dataset, and the ensemble model demonstrated the highest overall accuracy. Based on the best-performing model, critical rainfall threshold curves corresponding to 50%, 75%, and 90% levels of landslide susceptibility were established. This approach resulted in a risk-based threshold framework suitable for practical early warning applications. According to the results, all models achieved an overall accuracy exceeding 85%, and all adopted conservative classification strategies, reducing the risk of false negatives in real-world scenarios. In the case of Typhoon Morakot, the proposed rainfall thresholds successfully issued warning signals up to ten hours before the onset of landslides. For Typhoon Lupit, it was observed that the rainfall pattern easily influences the threshold curve. Therefore, using the average rainfall intensity and cumulative rainfall threshold values to cross-validate each other can help reduce the risk of false alarms.