以地理資訊系統搭配羅吉斯回歸、不安定指數及支撐向量機建立山崩潛感之比較分析

A GIS-based Comparative Study of the use of A Logistic Regression, the Instability Index Method and A Support Vector Machine for Landslide Susceptibility Analysis

詹勳全(Hsun-Chuan Chan) ； 陳柏安(Po-An Chen) ； 溫祐霆(Yu-Ting Wen)

《中華水土保持學報》 46卷4期 (2015/12) Pp. 213-222

https://doi.org/10.29417/JCSWC.201512_46(4).0003

摘要

本研究針對景山溪上游集水區，利用羅吉斯回歸、不安定指數及支撐向量機，搭配經濟部中央地質調查所製作的崩塌目錄，先初選10個潛在因子，再經過因子複選過程選取6個解釋山崩能力較佳的因子，建立模型並繪製山崩潛感圖，分析結果以ROC曲線（Receiver Operating Characteristic curve）評估三種模型預估山崩潛感之準確程度。研究結果顯示，地形粗糙度因子於羅吉斯回歸及不安定指數分析方法中，對山崩潛感值有一定的影響程度；而不安定指數不僅在河道附近之潛勢會有低估之情形，針對影響山崩潛感權重較大之因子，其分級之級數多寡亦會左右潛感分析結果；支撐向量機利用提供的資料在進行分類，其山崩潛感不會發生傾向於某個因子的情況，因此不會有因子權重擴大或縮小而影響模式判別的情況產生。另外，本研究將山崩潛感值分成低潛感、中潛感、中高潛感及高潛感四種情況，在羅吉斯回歸及支撐向量機分析中，實際山崩區域多座落於中高潛感以上之區域，由ROC曲線下面積（Area Under the Curve, AUC）得知，支撐向量機較其他兩種模式有較高的精確度，其AUC值為0.825優於羅吉斯回歸AUC值0.721及不安定指數AUC值0.718，在研究區內支撐向量機之山崩潛感分析結果較其他兩種模式優越。

關鍵字

羅吉斯回歸 ； 不安定指數 ； 支撐向量機 ； 山崩潛感

並列摘要

This study uses the inventories for landslides that were established by the Central Geological Survey for landslide data. A logistic regression, the Instability index method and a Support vector machine (SVM) are used to establish landslide susceptibility models and to produce landslide susceptibility maps for the upstream areas of Jing-Shan River. Ten causative factors for landslides are selected similarly to previous studies. A selection procedure is then used to reduce the number of factors. The receiver operating characteristic curve is used to evaluate the accuracy of the model results. The Logistic regression and the Instability index method both show that the roughness of the terrain is a critical factor for the susceptibility value. The instability index method can lead to possible underestimation around the river and the number of factor classifications can impact the results. SVM establishes the model by classifying the landslide data. The landslide susceptibility values are not reliant on particular factors. Therefore, the results for the model prediction are not influenced by the weights of the factors. The landslide susceptibilities are classified into four groups: low, medium, medium-high and high. SVM and Logistic regression are superior to the Instability index method because they identify landslides in medium-high and high susceptibility areas. The analysis of the area under the curve (AUC) gives an AUC value of 0.825 using SVM, 0.721 using logistic regression and 0.718 using the instability index method. This demonstrates that SVM is the best method to assess landslide risk in the research areas.

並列關鍵字

Logistic regression ； instability index ； support vector machine ； landslide susceptibility