台灣地區因山區地形陡峻,近年來,平原地區之開發已趨飽和,坡地之開發利用已是必然趨勢,全球極端氣候的影響,每逢颱風或暴雨來襲,就容易因集中性降雨而引發崩塌現象,並造成土砂災害,而土砂災害伴隨著嚴重的生命財產損失,因此建立災害損失評估模式,提供政府防災規劃時之參考實屬必要。 本研究比較三種衛星影像分類判釋方式(包括傳統之最大概似法、類神經網路、基因演算自動演化 類神經網路),以莫拉克颱風侵襲台灣期間,南台灣集水區為研究範圍,藉以獲取災害紀錄資料,並以地理資訊系統建置研究區屬性與空間資料庫。本研究參考前人之研究擇取之災害潛勢因子包括「坡度」、「坡向」、「地質」、「高程」、「距水系距離」、「距斷層距離」、「地形粗糙度」、「坡度粗糙度」、「曲率」、「坡地利用」與「有效累積雨量」等因子,研究中各災害影響因子透過相關性檢定,檢定各潛勢因子間之相關程度,結合判釋精確度最佳結果之基因演算最佳化類神經網路技術,透過多變量不安定指數建置山崩災害潛勢模式,並採用對數常態分配,推算研究區土砂災害潛勢機率,利用地理資訊系統繪製區域土砂災害潛勢機率圖。 本研究藉由現地調查,及政府所公布之災害資料與相關文獻整理後,推估研究區域內之財產損失,配合多變量不安定指數分析結果之土砂災害潛勢機率,推算建物、農地、林地及交通水利用地之災害損失金額,統合發展災害損失風險評估模式。本研究並藉由建置土砂災害損失評估模式,結合土砂災害潛勢機率圖,繪製研究地區相對應之災害損失推估。研究結果將可提供政府或相關防救災單位針對高土砂災害潛勢地區擬訂治理對策之參考,期可減少人民生命財產之損失。
Recently, the development in the slope land is inevitable in Taiwan since the densely populated plain areas are highly developed. With more and more concentrated extreme rainfall events as a result of climate change, in Taiwan, mass cover soil erosion occurs frequently and leads to sediment disaster in high and precipitous region during typhoons or torrential rain storms, causing a lost to the property and even the casualty of the residents in the affected areas. To make the execution of the regulation of slope land development more efficiency, construction of evaluation model for disaster loss is very important. In this research, traditional maximum likelihood method, genetic algorithm, genetic adaptive neural networks are employed for the classification of satellite images of watershed of southern Taiwan after 2009 Typhoon Morakot to obtain digital records of the ground and disaster. The slope, aspect, geology, elevation, distance to a drainage, distance to a fault, terrain roughness, slope roughness, curvature, level of development, and effective accumulative rainfall are used as fundamental factors in this research since they are used mostly in the references for their notable influence to landslide disasters and they are easy to acquired. Before inputting the above potential factors into evaluation model, correlation analysis must be done to avoid linear relationship between factors. By employing geographic information system ArcGIS, multivartive dangerous value method with lognormal distribution, we use genetic adaptive neural network to classify the satellite image in research area and to plot the sediment disaster potential map. This study employs field investigation, official disaster record, related references, and the probability predicted by multivartive dangerous value method to estimate the disaster loss for building, farm land, forestry land and land used for communication and water conservancy. The proposed evaluation model of disaster loss can be an effective tool for property loss estimation and the proper countermeasures to prevent serious damages in high potential regions from sediment disaster can be made.