地文因子對崩塌發生機率之影響-以高屏溪流域為例

Abstract

本研究主要係以高屏溪流域集水區為例，探討地文因子對於崩塌發生機率之影響。文中先將前人研究較常使用之地文因子以無母數統計檢定其與崩塌發生之相關性及因子間之獨立性，並結合物理機制篩選出坡度、順向坡比及崩塌率等3地文因子與崩塌發生具顯著相關，其中包含具時變性之因子(即崩塌率)，再結合水文因子(累積降雨量)以邏輯斯迴歸建立崩塌發生機率評估模式，即崩塌發生機率隨坡度、順向坡比、崩塌率及累積雨量增加而升高。該模式於訓練階段及驗證階段之整體正確率均約81%；研究中另以20次隨機挑選之歷史降雨事件進行驗證，包含崩塌發生與不發生之事件各10次，結果顯示倘設定模式P=0.5做為警戒門檻機率，模式判定之正確率約80%，與模式建置訓練階段及驗證階段之整體正確率相符合，僅2次誤警示及2次未預測崩塌發生。 本文進一步採用支持向量機(Support Vector Machine)、最小平方支持向量機(Least Square Support Vector Machine)與關聯向量機(Relevance Vector Machine)進行推估，並與邏輯斯迴歸結果相較，SVM將推估準確度從81.3％提高到82.44％，LS-SVM進一步將準確度提高到84.7％，且任何採用正規化方法之RVM均優於邏輯斯回歸模式。 本研究所發展之崩塌發生機率評估模式可應用於坡地之防災：其一，無論係以工程手段或植生工法減少一區域之崩塌面積，即可有效降低該區域之崩塌再發性。其二，一地區倘歷經地震、颱風或人為開發等大規模改變事件後，可參考本研究所發展之模式更新地文因子，校正崩塌發生之降雨門檻值。其三，可給定一適當之累積雨量及可接受之崩塌發生風險，則一集水區內需優先進行崩塌防治之區域亦可藉此評估而得。

This study investigated the influence of physiographic factors upon the probability of slide occurrence in the Kaoping River Basin. According to literature, statistical tests and physical mechanisms, three physiographic factors (slope steepness, the dip slope ratio, and the time-dependent landslide ratio) were significantly related to slide occurrence. Together with a hydrological factor (i.e., cumulative rainfall) to establish an assessment model for estimating the probability of slide occurrence using logistic regression and machine learning. The model’s overall accuracy in the training and validation stages was about 81% and 87%, respectively. Twenty randomly selected historical rainfall events were employed for verification, half of them are with slide occurrence. The results showed that the model accuracy was approximately 80%, if the probability threshold Psh is set to be 0.5. In addition to logistic learning, the machine learning algorithm such as Support Vector Machine, Least-Square Support Vector Machine and Relevance Vector Machine are used to estimate the probability of slide occurrence, and compared with the logistic regression results. The SVM enhances the estimation accuracy from 81.3% to 82.44%, the LS-SVM further increases the accuracy to 84.7%. The proposed RVM with any kind of normalization is superior to those of the logistic model. This assessment model can assist in prediction of slide occurrence and the subsequent/corresponding engineering measures or vegetation restoration can be utilized to reduce the landslide ratio and the probability of slide occurrence. After landscape changes, the model’s physiographic factors can be updated to adjust the rainfall threshold for slide occurrence. Given a cumulative rainfall and an acceptable risk of slide occurrence, the proposed model can identify priority regions for slide prevention.