台灣地區因破碎的地質構造，常因降雨、地震等觸發山崩造成災害。近年來因遙測技術以及地理資訊系統的蓬勃發展，山崩的分析已有不錯的監控與評估成果，並且累積了豐富的研究資料。為了直接利用、整合眾多不同來源的調查資料，找出隱含在資料中的知識，以利下一步運用。有許多針對具空間屬性資料的分析方法，能從大量的山崩調查資料中萃取出幫助辨識崩塌的有效法則。 本研究蒐集2004年至2008年颱風季節時的石門水庫集水區崩塌資料以及相關遙測及空間資料，主要有高解析度衛星影像、DTM資料、土地利用類別、地文描述資料(如河流、斷層)等。透過資料前處理與轉換，結合資料挖掘的學習法，找出山崩與環境因子間的關聯，建立降雨觸發的山崩特性推論模型，了解影響山崩的重要因子。而決策與規則訓練模型與測試成果，可調整資料以提升辨識精度。本研究所建立之決策樹針對2004年艾利颱風於研究區內之測試資料進行自動化崩塌地辨識，精度可提升至79%。本研究成果顯示，以資料挖掘技術對於遙測影像及GIS資料進行颱風豪雨誘發之崩塌地辨識為可行的方法。 The fractural geological conditions in Taiwan have caused serious landslides in mountainous regions after typhoon or earthquake every year. Remote sensing and other spatial data have been used successfully to evaluate and monitor landslide hazards. Satellite remote sensing and GIS-based data are effective sources to obtain information about environmental conditions covering large areas with high spatial details. For landslide related issues, the effect of environmental characteristics on the probability of landslide is an important factor and commonly used to predict landslide risks. In addition, other spatial data, such as digital terrainn model (DTM), land-cover types, vegetation, soil, and other natural and man-made factors may all contribute to the prediction of landslide susceptibility. This study utilizes data mining techniques to analyze complicated datasets in order to understand landslide risks in the Shihmen Reservoir watershed located in northern Taiwan. An inventory of collected known landslides caused by typhoons from 2004 to 2007 in the study site is used as training data. Decision rules for detecting landslide from selected attributes have been established. The rules are applied to predict landslides induced by typhoons. The rules constructed from decision tree algorithms are refined to improve the classification accuracy. The identification accuracy is about 79% for the test data with 2004 Aere typhoon. With the developed algorithms and data mining techniques, landslides induced by heavy rainfall can be mapped efficiently from remotely sensed images and geo-spatial analysis.
