本研究沿用地調所山崩潛感分析計畫之分析模式及各項原始資料，進行各種驗證分析，並嘗試進行部分重要因子處理之精緻化研究及各個因子權重穩定性的探討。同時比較因子在分區評分與否暨內部評分與否的成效，以評估山崩潛感分析計劃分析模式之各種優劣點及探討可能改善的空間。本研究不僅利用不同統計軟體驗證本研究所發展之程式，同時也比較由羅吉斯迴歸及模糊類神經網路所進行的山崩潛感分析之成果，希望能夠暸解目前最常被使用在山崩潛感分析的三種方法之間的差異性。 本研究針對坡度及地形粗糙度兩項因子進行深入研究，包括：（1）使用高通濾波的方式粹取地表細部起伏的變化，並藉此地形製作新的地形粗糙度；（2）使用使用韋伯分布的累積密度函數針對各事件坡度崩壞比曲線進行擬合及內部評分。本研究並將其他項重要因子重新處理及重新迴歸崩壞比曲線，最後針對山崩與非山崩組作亂數選取相近的格網數作為樣本加入判別分析進行山崩潛感分析。 將精緻化及修正後之因子加入判別分析的結果，不僅降低了潛感因子間的相依性，且得到之山崩組準確率比地調所山崩潛感分析計畫之準確率顯著提升。經不同亂數取樣後所得到各個因子的權重，其標準差皆相當的小，顯示因子之權重是相當穩定的。判別分析、羅吉斯迴歸與模糊類神經網路相互比較分析成果後，發現三種分析方法在準確率上相差不大，表示判別分析在對於山崩潛感分析的預測模式上，並不會輸給需要長時間訓練的模糊類神經網路方法。 因子有無按崩壞比做內部評分的動作，對於準確率的影響並不十分顯著，但以崩壞比做因子內部評分，去除了門檻值以下影響不大的資料，較能彰顯該因子重要值域的影響，分析結果更能有效地表現出此因子對於判別模式的貢獻。在以地域單元分別做因子內部評分與否的問題上，不論是有分區或是無分區的情形下，因子皆會被正規化至0 ~ 1之間，其中各分區崩壞比迴歸式的斜率差，並不會對各分區之內部評分造成太大的差異，最後的分析結果也影響不大。 This study follows the methodology and uses the original data of a landslide susceptibility project of Central Geological Survey, Taiwan (CGS). This study proceeds to check the data and to validate the model, and improves the treatment of some of the important factors. Reliability of weights among the factors was tested. The necessity of internal rating of each factor according to a terrain unit was also tested, and possible improvement was discussed. This study used different statistical software to validate the program we developed. I also compared the results evaluated by the logistic regression and the fuzzy neural network method so that the superiority among the three frequently used methods in landslide susceptibility analysis could be compared. Slope factor and terrain roughness factor were further studied. It includes : （1）Using high pass filter treatment to emphasize the local roughness of a terrain. （2）Using cumulative Weibull distribution to fit the curve of landslide ratio of slope factor. All factors were reproduced and redefine the internal rating of each factor were redefined. Samples for analysis were done by random sampling method from the non-landslide group so that they have approximately same number as the samples from landslide group. After the reprocessing and refinement of the factors, the result for each different event is significantly improved. Different random sampling results provide different weights. The result shows that the standard deviation of a weight for each factor is small and means the weights are stable and reliable. The results among the discriminant analysis, the logistic regression and the fuzzy neural network are comparable in overall accuracy. This indicates that the result from discriminant analysis is as good as the fuzzy neural network method which takes much time to train the sample. Internal rating of a factor according to the landslide ratio doesn’t affect the accuracy very much, but if the factor is rated according to landslide ratio can minimize the effect of data which are out of lower threshold or higher threshold, and emphasize the effect of the important range of the factor, and make a factor more effective in discriminant analysis. Whatever the internal rating of a factor is based on terrain units or not, the score of each factor will be normalized to a range between 0 and 1, and the result is not significantly different.
