台灣位在西北太平洋颱風移行的主要路徑上，依據統計平均每年約有 3 至 4 個颱風可能侵襲台灣地區，伴隨颱風而來的強風豪雨所引發的洪水與坡地土砂崩塌災害，不僅嚴重威脅人民的生命安全，更時常造成社會經濟的重大損失。過去崩塌災情分析方式，常由經驗豐富地質專家進行人工研判，不僅耗時、費力，無法在短時間內做出有效的災害評估與災後復原計畫；且易因高山阻隔或災後道路中斷等因素，使得大範圍災情調查工作無法順利進行。因此，如何提升自動化判別災區崩塌地區域之效能，是一個值得探討的研究課題。 本研究使用高解析衛星影像與地形資料，以提升衛星影像自動化暴雨型崩塌地判釋之精度為目標，除了使用在許多影像分類技術常被採用的最大似然法及倒傳遞類神經網路法外，亦建立以支持向量機法為基礎之崩塌地判釋機制；文中除了對三種方法進行分類精度評估，亦探討影像色調及地形特徵作為暴雨型崩塌地判釋特徵之顯著性，以強化崩塌地分析的自動化及效能。 從分離度分析結果可顯示訓練樣本之選取適當，坡度特徵的加入有助於增加崩塌地與其他三種類別分離度。此外，由分類精度評估結果可知 BPNN法與 SVM法分類精度均優於 ML法；至於判釋特徵，如坡度、OHM 等之使用上，對於山崩判釋正確率提升上有不同程度之影響，其中以坡度特徵影響最為顯著。然而，因混合像元效應 (本次實驗資料解析度 20m)、取樣率不一之影響，對於本研究使用特徵及分類方法，值得進一步探討與分析。

Taiwan is located in the northwest Pacific, in a major migration path of typhoons. There are about 3-4 typhoons likely to hit the Taiwan area in average annually. Strong winds and heavy rains are accompanied with typhoons, thus, usually leading to a result of floods and landslide hazards. Not only human lives are seriously threatened, but also social and economic losses might be imposed. It is estimated that the economic loss is about six billion US dollars in every year due to typhoon. Landslide inventory used to be done by manual judgement of experienced geologists. It is a time consuming and labour-intensive job, to lead making effective disaster assessment and recovery plans impossible. And let large-scale disaster investigation can not proceed smoothly due to blocked by mountains or post-disaster factors such as road disruption. Therefore, how to enhance the automation and its performance of landslide inventory is an important research topic. Multi-source high-resolution data, e.g. a SPOT satellite image, 5m x 5m DTM reduced from a LIDAR data and aerial orthophotos, are fused to construct the feature space for landslides analysis in this paper. Then, those spectral and geomorphometric features are used to recognize landslides by a Maximum Likelihood (ML) method, an Artificial Neural Network (ANN) method and a Supported Vector Machine (SVM) method. The classification results are evaluated in comparison with those of manual-interpretation. Moreover, the separability analysis for the used features on rainfall-induced landslide interpretation is also provided. The separability analysis result indicates that slope is an important factor to distinguish the landslide and other classes. In this case study, the recognition accuracy for landslides and non-landslides for the BPNN and SVM method are better than the ones for the ML method. Moreover, slope is a significant interpretation key for landslides recognition. Due to mixed pixel effect (resolution of this experimental data is 20m) and the effect of different sampling rates, the characteristics and assessment for the conducted methods is worth further study and analysis.