| 摘要: | 目前山崩潛感分析以統計方法為主流，其中以多變量分析為最常使用的方法。本研究嘗試以多變量分析中的羅吉斯迴歸方法進行山崩潛感分析的研究，並繪製山崩潛感圖。 研究區位在臺灣中北部的國姓地區，為地調所五萬分之一地質圖國姓幅的範圍。山崩目錄由集集地震、賀伯、桃芝及敏督利颱風共四次事件前後，八幅不同時期的SPOT衛星影像判釋及數化崩塌地而得。數值地形資料的原始解析度為40公尺，經曲面內插為20公尺來使用。使用的潛在因子有岩性、坡度、坡向、坡度粗糙度、總曲率、植生植數、全坡高等；促崩因子則使用愛氏震度及最大時雨量。 各因子經地理資訊系統處理成格網式資料，分為山崩組與非山崩組。研究區以雙冬斷層為界，分為東邊之高山區及西邊之淺山區，分區進行各因子崩壞比統計，接著找出崩壞比迴歸式，將因子值轉換為崩壞比值，最後將崩壞比值正規化至0至1，做為因子內部評分之分數。對於山崩組及非山崩組，以亂數取樣的方式選取相同的格網數目的資料，輸入SPSS統計軟體進行羅吉斯迴歸分析並建置迴歸模型。全區資料代入羅吉斯迴歸模型計算出各格網之山崩潛感值，將潛感值劃分為高潛感、中高潛感、中潛感及低潛感四個等級，繪製山崩潛感圖。 分析結果顯示，四次事件之總體準確率結果從72.5%至90.6%，且經套疊實際山崩比對，山崩大部分位於高潛感區，預測成效尚稱滿意。山崩潛感圖的高潛感區表示未來遭遇相似強度的地震或豪雨事件，極易再度發生山崩的地點，各點的山崩機率並可由事先建立的機率模式計算而得。 Statistical methods are the main stream in landslide susceptibility analysis recently. Multivariate analysis is the most popular method among those statistical methods. The purpose of this study is to assess landslide susceptibility by Logistic regression, one of the multivariate analysis methods, and to examine the performance of this approach.. The study area locates at Kouhsing area, in central Taiwan. This area is taken as the same area as the 1:50000 geological map of Kouhsing sheet of CGS (Central Geological Survey). We collect eight SPOT satellite imageries covering four triggering events, including the Chi-Chi Earthquake, the Herb Typhoon, the Toraji Typhoon and the Mindule Typhoon. Using those pre-event SPOT imageries and after-event SPOT imageries to derive the landslide inventories. The 40m×40m resolution digital terrain model was inserted to 20m grids. Lithology, slope, slope aspect, terrain roughness, slope roughness, terrain curvature, NDVI (normalized difference vegetation index) and total slope high are taken as potential factors. Arias intensity and maximum hourly rainfall are taken as trigger factors. The raster cell data extract from every factor, divided into landslide group and non-landslide group. Furthermore, this study area is divided into hill zone on west part and mountainous zone on east part. The two zones are separated by the Shuantung Fault. We calculate the propotion of failure for every factor on each zone. We find a fit line of the propotion of failure on each zone, then convert the factor value to propotion of failure. Finally, we scale the propotion of failure to a score that ranges between 0 to 1. We sample equal cell number of data randomly for landslide group and non-landslide group, then input those data to SPSS statistical software and build a logistic model for the study area. We then apply the model to the whole study area, and calculate a landslide susceptibility index for each cell. We further divide the susceptibility index into high, moderately-high, medium and low classes to produce a landslide susceptibility map. Overall accuracy in these four events are 72.5% to 90.6%. Large parts of landslide that locate in high susceptibility area indicate the results are satisfactory. The cells that locate in high susceptibility area indicate the cells may have slid during the similar earthquake event or typhoon event in the future. |
