運用衛星影像進行六龜區域崩塌地景監測與影響因素分析

Monitoring and Susceptibility Assessment of Landslides in the Liouguei Area Using Satellite Imagery

本研究目的在於運用多時期衛星影像，結合物件導向（Object-Based）方法和支持向量機（support vector machine, SVM）分類技術，以及地景指標，對六龜試驗林莫拉克颱風前至2024年間的崩塌地變動進行監測，並進一步利用邏輯迴歸進行崩塌潛勢預測。結果顯示，崩塌地面積在2009年8月由111.90 ha急劇增加至748.00 ha，隨後逐漸減少至2024年的230.46 ha。而地景破碎化和複雜度在災後也隨時間慢慢減少，顯示地景結構逐漸恢復完整連續。透過新增崩塌地分析的結果，在2011至2024年間，崩塌地面積變化存在波動，其中2013年達到顯著峰值，新增崩塌地面積達253 ha，與2013年發生多起重大豪雨事件相符。在邏輯迴歸模型分析結果，顯示高程、坡度、距河川距離和坡向為主要影響崩塌潛勢的因素，其中以河川距離和坡度對崩塌發生的影響最為顯著，模型效能評估顯示，AUC值為0.704，總體準確率為64.2%，這結果也代表著模型在崩塌潛勢預測具備一定的參考價值。總體而言，本研究利用人工智慧分類方法，量化時間序列崩塌資料，不僅提供了崩塌地長期監測的資訊和技術參考，也為區域防災減災策略制定提供了重要的科學依據。

The purpose of this study is to utilize multi-temporal satellite imagery, combined with Object-Based methods and Support Vector Machine (SVM) classification techniques, along with landscape metrics, to monitor landslide dynamics in the Liouguei Experimental Forest from before Typhoon Morakot up until 2024. Additionally, the study aims to further apply logistic regression for predicting landslide susceptibility. The results showed that the landslide area increased sharply from 111.9 hectares in August 2009 to 748 hectares, then gradually decreased to 230.46 hectares by 2024. The landscape fragmentation and complexity also gradually decreased over time after the disaster, indicating that the landscape structure is slowly recovering to a more intact and continuous state. The analysis of new landslides between 2011 and 2024 revealed fluctuations, with a significant peak in 2013, when the new landslide area reached 253 hectares, corresponding to several major heavy rainfall events that year. Logistic regression analysis identified elevation, slope, distance to rivers, and aspect as key factors influencing landslide susceptibility, with distance to rivers and slope having the most significant impact. Model performance evaluation showed an AUC value of 0.704 and an overall accuracy of 64.2%, indicating that the model has some reference value for landslide susceptibility prediction. Overall, this study demonstrates the use of AI-based classification methods to quantify time-series landslide data, providing valuable information and technical references for long-term landslide monitoring and offering critical scientific support for regional disaster prevention and mitigation strategies.