以多個潛感因子進行機器學習演算法之邊坡滑動潛勢分析

Abstract

坡地災害是臺灣常見的災害之一，崩塌的發生不僅可能造成財產損失，更可能危及人命。因此，瞭解山坡地發生崩塌的潛勢機率成為大地防災工程中不可或缺的資訊，透過事先獲取各山坡地之崩塌潛勢值，在天災如颱風、豪雨來臨前提供工程師即時資訊，事先於高潛勢坡地進行邊坡穩定工程。本研究旨在研發一套機器學習崩塌潛勢系統模型，透過輸入一系列崩塌潛感因子，將雙北市境內的山坡地給予初步之崩塌潛感值，透過該潛感值，便能事先於高潛感之山坡地進行對應之邊坡穩定工法，藉以降低未來發生崩塌災害之機率。

本研究蒐集15項崩塌潛感因子，包含坡度、坡向、高程、剖面曲率、地形粗糙度、植生指標、建築物面積百分比、水系距、道路距、斷層距、順向坡指標、褶皺指標、地質敏感區指標、InSAR變位指標與雨量指標與其對應之地真資料，以隨機森林以及極限梯度提升兩種機器學習演算法對臺北市以及新北市境內之山坡地進行崩塌潛勢分析，透過蒐集蘇迪勒颱風侵臺期間所造成之上述資料進行機器學習模型之訓練。潛感因子中，依降雨型態之差別將輸入的潛感因子分為兩類型，第一類型為14項潛感因子加累積月雨量，第二類型為14項潛感因子加I3及R24，依此方法總共建立了四種機器學習模型，並以其他未納入訓練之崩塌地和雙北歷史上之著名山崩事件進行模型驗證。

本研究結果顯示，以蘇迪勒颱風崩塌之70%資料所建立起之模型，能有效預測出其餘30%之測試單元崩塌或不崩塌的結果，尤其以XGBoost模型表現更為優異而其AUROC值均超過0.9，顯示該模型確有辨別崩塌與否之能力。並於未納入訓練之22個蘇迪勒颱風崩塌斜坡單元所展現出的預測結果優異，只有1組結果顯示為預測不崩塌，其餘43組結果都能成功預測到崩塌。於臺鐵瑞芳-猴硐路段崩塌之驗證表現良好，平均崩塌機率為59.3%。此外，累積月雨量、I3、R24、地形粗糙度、坡度及高程對模型之成果有高度影響，其因子之重要度為眾多因子中數值偏高者，顯示山崩需要在足夠條件之地形特徵以及雨量促崩下發生。

Landslides are one of the common disasters in Taiwan, resulting in not only property loss but also casualties. Therefore, understanding the potential for landslides in mountainous areas is crucial information in Geo-disaster engineering. By obtaining the preliminary landslide susceptibility values for each slope, engineers can be provided with timely information before natural disasters such as typhoons and heavy rains, allowing them to implement slope stabilization measures in high-susceptibility areas in advance. The objective of this study is to develop a machine learning model for landslide susceptibility assessment. By inputting a series of landslide susceptibility factors, the model can provide initial landslide susceptibility values for slopes within Taipei and New Taipei City. Based on these values, corresponding slope stabilization techniques can be applied in high-susceptibility areas, thereby reducing the probability of future landslide disasters.

Two machine learning algorithms, namely Random Forest and Extreme Gradient Boosting, were established using 15 landslide susceptibility factors in this study. These models were trained by inputting failure and non-failure slope units that occurred during the period of Typhoon Soudelor in 2015 and validated using landslide inventories within Taipei city.

The models established in the present study are capable of distinguishing between failure and non-failure slope units. High AUROC values (i.e., AUROC values higher than 0.9) were obtained for the four landslide susceptibility models. Twenty-two slope units from the Typhoon Soudelor event, which were excluded from the training data, were used as validation data, and 21 of them were successfully predicted. The models also performed well in validating the landslide occurrence along the Ruifang-Houtong section of the Taiwan Railways, with an average landslide probability of 59.3%. The landslide susceptibility factors of accumulated monthly rainfall, I3, R24, terrain roughness, slope degree, and elevation have a significant impact on the performance of the models, with the higher factor importance. This indicates that landslides are triggered by not only terrain conditions but also the impact of rainfall.