| 摘要: | 事件型山崩潛感模型是利用事件誘發山崩目錄及地形、地質因子和促崩因子訓練建模。若建立模型的事件之前有超大豪雨或颱風或大地震發生過，事件誘發山崩的分布特性可能與未受到干擾的情況有所不同。本研究訂定時隔6個月以上第一場超過降雨門檻的雨量事件為獨立事件，在曾文水庫集水區挑選賀伯颱風、桃芝颱風、敏督利颱風、20050615豪雨、20060609豪雨、莫拉克颱風、20110718豪雨、20120610豪雨、20150523豪雨共九期獨立事件建立山崩潛感模型。將模型進行交叉驗證，發現以豪雨事件建立的模型彼此之間相互預測結果表現不錯。敏督利模型由於坡度因子及促崩因子表現良好，預測其他事件的AUC皆達到0.710以上，解釋莫拉克颱風誘發山崩的AUC達0.673，為最穩定的事件型山崩潛感模型。
比較歷年山崩目錄按傳統方法建立之不包括促崩因子的山崩潛感模型與各個事件基礎山崩潛感圖，可以觀察到九個事件個別建立的基礎山崩潛感圖與歷年山崩潛感圖會有相近的潛感分布趨勢，表示歷年山崩潛感圖即可代表一個區域的基礎潛感値分布。本研究選用歷年山崩潛感模型為基礎潛感模型，將各事件分為極端事件及一般事件，將不同潛感區間中相同雨量條件下的崩壞比取其平均值，分別建立山崩潛感‒崩壞比和最大時雨量/總雨量關係式，利用雨量因子間的相關係數將關係式合併以計算在不同雨量分布下各個潛感體質的崩壞比。各事件山崩機率圖的預估山崩面積與事件誘發山崩目錄比較結果顯示，本研究建立之關係式低估桃芝颱風誘發山崩；高估賀伯颱風、敏督利颱風及豪雨事件誘發山崩，預測莫拉克颱風誘發山崩與實際誘發山崩面積較為吻合。;An event-based landslide susceptibility model is constructed based upon an event landslide inventory, topographic factors, geological factors and triggering factors. If an extreme rainfall, typhoon, or a major earthquake happened before a modeling event, the characteristics of landslide distribution may be different with undisturbed conditions. In the present study, an independent event is defined as the event without a prior-event which exceeds the rainfall threshold of the region within 6 months or more. In the Zengwen reservoir catchment area, this study chooses nine independent events, including Herb typhoon, Toraji typhoon, Mindulle typhoon, 20050615 rainfall, 20060609 rainfall, Morakot typhoon, 20110718 rainfall, 20120610 rainfall, and 20150523 rainfall, to establish nine event-based landslide susceptibility models. These models are then cross-validated. The results are good between rainfall events. Due to good performance of slope factor and triggering factors, Mindulle model is the most stable event-based landslide susceptibility model for typhoon events. It shows that AUC of the prediction curve for landslide induced by Morakot typhoon is 0.673; AUCs of prediction curve for other events are more than 0.710.
Compare the landslide susceptibility model built by multi-temporal landslide inventories, via traditional approach without triggering factors, to the nine basic susceptibility maps built by each event, a similar trend of susceptibility distribution among them can be observed. This study chooses the multi-temporal landslide susceptibility map as representative basic susceptibility of the region. The events are divieded into extreme events and common events. In each susceptibility bin and each rainfall bin, the average of probability of landslide failure is calculated from every events landslide inventory of same type, and then a relationship among landslide susceptibility, probability of failure, and rainfall intensity or total rainfall is completed. Utilizing the correlation coefficient between two rainfall factors, this study combines the two relationships to calculate the probability of failure of different susceptibility values in rainfall event. Comparing the result of landslide area predicted by the relationship and each event inventory shows that the relationship underestimate landslide area of Toraji typhoon, and overestimate landslide area of Herb typhoon, Mindulle typhoom and rainfall events. Predicted landslide area of Morakot typhoon is approximate to actual landslide area. |
