山崩潛感模型的逐步建構與決策分析－以陳有蘭集水區為例

Progressive Modeling and Decision Analysis of Landslide Susceptibility: Case Study in Chenyulan River Watershed, Taiwan

葉柏村(Po-Tsun Yeh) ； 張光宗(Kuang-Tsung Chang)

《中華水土保持學報》 50卷2期 (2019/06) Pp. 49-62

https://doi.org/10.29417/JCSWC.201906_50(2).0002

摘要

本研究針對陳有蘭集水區受2009年莫拉克颱風降雨誘發山崩事件，利用包含弱面位態資料的249個邊坡單元，並考慮雨量、地形、地質與邊坡狀態（最近是否曾經崩過）等相關因子，採用二項式邏輯斯迴歸方法建立山崩潛感的預測模型，並做適合度檢定，再以ROC曲線分析結合決策分析中的損失函數找出最佳截止點，利用最佳截止點所對應到的機率門檻值來計算模型準確率（77%）。整體而言，由於本研究模型的建構過程採逐步篩選各項因子的統計方式，使得模型可在使用較少的因子的情況下，也能有效預測陳有蘭溪集水區的山崩事件。此外，研究結果發現岩體的弱面位態與最近是否曾經崩過之因子，與山崩潛感有較顯著的關係，可供後續應用於此類山崩潛感模型的參考。

關鍵字

山崩潛感模型 ； 二項式邏輯斯迴歸 ； 損失函數 ； 弱面位態 ； 最近是否曾經崩過

並列摘要

Many rainfall-triggered landslides occurred in the Chenyulan river watershed in Central Taiwan during typhoon Morakot in 2009. In this study, the landslide susceptibility model of the Chenyulan river watershed was developed and evaluated using a binomial logistic regression model based on 249 slope units. The factors considered while developing the landslide susceptibility model were rainfall, topography, geology, and slope condition (i.e., presence or absence of previous landslide events). Next, we tested the goodness-of-fit of the regression. Moreover, we performed receiver operating characteristic curve analysis with loss function to determine the best cutoff point. The cutoff value of probability corresponding to the best cutoff point was used to calculate the accuracy of the model (77%). Overall, the model effectively predicted landslides by using relatively few factors because of progressive modeling and analysis. In addition, the results showed a significant relationship between landslide susceptibility and geological factors (the orientation of the weak plane) and the condition of slope (presence or absence of previous landslide events). The present study can provide a reference for subsequent studies on this topic.

並列關鍵字

Landslide susceptibility model ； binomial logistic regression ； loss function ； orientation of the weak plane ； presence or absence of previous landslide events