On the Rainfall Induced Landslide Susceptibility and the Control Factors of the Analysis Accuracy - at the Slate Formation Area in Central Taiwan

本研究針對台灣中部板岩區之山崩潛勢相關之因素做進一步的探討，利用羅吉斯回歸與支援向量機針對敏督利颱風事件，在不同尺度之研究區，以20*20M之網格、不同門檻值劃定之集水區坡面單元以及坡向坡面單元等不同尺度之單元，進行山崩潛勢分析，再根據其ROC曲線之AUC面積驗證之，歸納出因子對山崩影響程度，選擇合適尺度並建立該山崩潛感模型。進一步利用不同雨型分析以及大氣環流模式推估之未來降雨量，應用敏督利颱風建立之山崩潛感模型進行未來山崩潛勢之預測。 本研究以區域因子(道路距及水系距)、地質因子(斷層距)、地形因子(高程、坡度、坡向、順逆向指標)、植生因子(NDVI)及雨量因子(崩塌降雨指標)作為山崩潛感分析之潛在因子，以不同尺度分析單元進行山崩潛感分析，結果顯示以敏督利事件作為訓練，不管是利用羅吉斯回歸還是支援向量機進行山崩潛勢分析，其ROC曲線之AUC面積判識準確率達80%~90%，其中，支援向量機比羅吉斯回歸還要準確。將建立模型套入未來降雨量，可進行未來之山崩潛勢預測，並分析其災害及風險，應用於減災防災，或是作為政府機構進行國土規劃(包含土地管理、道路及邊坡整治)之有效工具。 以不同尺度分析單元之山崩潛勢分析結果顯示，其繪製時間上以網格單元最快速，其次是集水區坡面單元，坡向坡面單元最費時間；在分析時間上，網格單元由於分析數量很多。因此，不管是在建模還是在進行山崩潛勢預測網格單元是最花時間的；在分析準確度上，網格單元最準確，因為分析單元越小越能準確捕捉訓練所需之數據與數量；在山崩潛感圖表現上，集水區坡面單元和坡向坡面單元較能準確表現出與實際地形的一致性，而網格是顯現出研究區潛感值分佈趨勢，將其應用在防災上，若是使用坡面單元進行分析而得之山崩潛感圖，可顯示出遇到天災時該單元上的整個區域之危害度，從而做好防災減災的準備。 以不同尺度之研究範圍進行山崩潛勢分析之結果顯示，選定小範圍個別流域的研究區所得到之潛感值，相對於大範圍三流域研究區之潛感值有判識偏保守之可能，以大甲溪流域為例，該研究區在敏督利事件的山崩目錄之崩塌地都在非板岩地區，但是本研究選擇大甲溪板岩區作為研究範圍，其訓練數據集中在中游且位於斷層上，而大甲溪流域之斷層都集中在板岩地區，所以在大甲溪板岩區進行山崩潛勢分析，其預測之潛勢值會偏高。This study explores the causative factors related to the landslide susceptibility in the slate area of central Taiwan. Applying the logistic regression and support vector machines, this study develop the landslide susceptibility models based on the landslide inventory of the 2004 Mindulle typhoon in different spatial scales, including the 20*20M grid, slope unit derived by aspect, and slope unit derived by elevation. According to the AUC values of the ROC curves, the accuracy of the landslide models was summarized; the appropriate scale is suggested and the landslide susceptibility models were established. This study further introduced the different rainfalls estimated by the rain pattern analysis and the general circulation model of TCCIP to the landslide susceptibility models to predict the landslide susceptibility in the future. In this study, considering the regional factors (distance to road and distance to river), geological factors (distance to fault), topographic factors (elevation, slope, aspect, and Ids), vegetation factor (NDVI), and rainfall factor (Id) as control factors, the landslide susceptibility was analyzed in different scales. The results show that, for the 2004 Mindulle typhoon, the AUC values of the logistic regression (LR) or support vector machines (SVM) models are in the range of 80%~90%. However, the performance of SVM is slightly better than LR. The models were then applied for the landslide susceptibility preditions. The models and the results could be used for disaster reduction and prevention, or for the government agencies in suburban management (including land use management, road maintenance, and slope remediation). The analysis results of the landslide susceptibility in of different scales show that the grid unit is the most efficient in mesh preparation, followed by the slope unit derived by elevation, and the slope unit derived by aspect. For the analysis time, however, the grid unit mesh consumes more time in the analyses. In terms of analysis accuracy, the grid unit could be the most accurate, because the smaller the analysis unit the better it can be trained. In the landslide susceptibility map, the results of slope unit can accurately represent the actual terrain, and the results of grid unit possess a trasition trend of landslide susceptibility in the study area. Applying it to disaster prevention, if we used the slope unit to analyze and obtain the landslide susceptibility map, it can better show the hazard level in the whole area, so as to prepare for disaster prevention and mitigation plans. The results of landslide susceptibility analysis in different scales indicate that landslide susceptibility obtained from the study areas of small scale individual watersheds are more conservative than the landslide susceptibility of the large scale area of three watersheds. Taking the Tachia River as an example, the landslide inventory of the Mindulle event in the study area are mainly in the non-slate area, but this study selected the slate formation area of the Tachia Watershed as the research area. The training data is concentrated in the middle reach and located near the faults. The faults of the Tachia watershed are concentrated in the slate area, so the landslide susceptibility in the Tachia slate area, predicted will be overestimated.摘要 i Abstract iv 圖目錄 ix 表目錄 xv 第一章 緒論 1 1.1研究動機與目的 1 1.2研究流程 1 1.3論文架構 2 第二章 文獻回顧 4 2.1山崩潛感分析因子決定相關之研究 4 2.2山崩潛感分析相關之研究 4 2.2.1羅吉斯回歸應用於山崩潛感分析 4 2.2.2支援向量機應用於山崩潛感分析 5 2.3預測雨量和極端氣候相關之研究 7 2.4山崩分析單元相關之研究 8 第三章 研究區域概述與研究方法 13 3.1研究區域概述 13 3.1.1大甲溪流域 13 3.1.2烏溪流域 14 3.1.3濁水溪流域 14 3.2機器學習方法分析與驗證方法 15 3.2.1羅吉斯回歸 15 3.2.2支援向量機 16 3.2.3 ROC曲線 18 第四章 山崩潛勢模型建立之關鍵問題 23 4.1誘發因子的分級和正規化 23 4.1.1因子的分級與選擇 23 4.1.2因子正規化 26 4.2流域範圍尺度效應 27 4.2.1山崩潛勢模型之分析流程 28 4.2.2羅吉斯回歸之模型建立 28 4.2.3支援向量機之模型建立 29 4.2.4結果與比較 29 4.3未來雨量之預測 31 4.3.1雨量站資料準備與整理 32 4.3.2降雨統計機率分佈 33 4.3.3不同重現期距之降雨量 34 4.3.4大氣環流模式推估未來降雨量 34 第五章 坡面單元之劃分與建立 67 5.1集水區坡面單元之劃分與建立方法 67 5.1.1集水區坡面單元之劃分 67 5.1.2集水區坡面單元之羅吉斯回歸模型建立 68 5.1.3集水區坡面單元之支援向量機模型建立 69 5.2坡向坡面單元之劃分與建立方法 70 5.2.1坡向坡面單元之劃分 70 5.2.2坡向坡面單元之羅吉斯回歸模型建立 71 5.2.3坡向坡面單元之支援向量機模型建立 72 5.3單元的選擇與分析結果 72 第六章 山崩潛勢模型分析結果與比較 87 6.1網格單元之羅吉斯回歸分析模式預測 87 6.2網格單元之支援向量機分析模式預測 92 6.3集水區坡面單元及坡向坡面單元分析模式預測 97 6.3.1集水區坡面單元分析結果 97 6.3.2坡向坡面單元分析結果 100 6.4結果與討論 102 第七章 結論與建議 134 7.1結論 134 7.2建議 135 參考文獻 136 附錄 野外調查之資料及其點位 13