在台灣，因多山的地形、頻繁的地震、暴雨及颱風的侵襲，導致崩塌災害經常發生。因此建立一個有效、並能產出合理精度成果的崩塌偵測方法顯得相當重要。物件式影像分析法為一新的影像分析技術，首先為依據影像中物件的均質性進行分割，之後則針對分割好的影像物件進行分析。不同於傳統的影像分析方法，影像切割能有效地整合各類資料並產製更精確的結果。本研究在方法上嘗試(1)偵測並繪製2009年莫拉克颱風所發生的崩塌地；(2)將已知對崩塌發生具有直接影響的地形參數 (包含坡度、坡向、高程、曲率與凸性) 透過於物件式影像分析法進行整合；(3) 透過比較研究展現此方法在崩塌偵測成果上的優勢，主要以總體精度 (overall accuracy)以及Kappa統計值檢驗之。為此，本研究以台灣南部高屏溪內花果山盆地作為研究區，並使用8公尺地面解析度的福衛二號多光譜影像(可見光以及近紅外光)，並整合10公尺地面解析度之數值高程模型所產製的地形參數進行研究。為了證明地形參數在崩塌偵測上的有效性，本研究比較了加入地形參數以及未加入地形參數的影像分類成果，具體來說，本研究針對6種不同的地表型態進行影像分類 – 裸露地、河道、森林、崩塌堆積區與崩塌侵蝕區，並使用隨機抽樣來評估分類精度，結果顯示當加入地型屬性資料，總體精度為81.8%，Kappa為0.64；沒有加入地形屬性資料的總體精度則為77.5%，Kappa為0. 55。從本研究的案例發現，加入地形參數後可降低崩塌的在陰影森林地區中所發生的分類錯誤，並可將建物、崩塌地以及裸露進行較好的區分。總體來說，研究成果顯示了以物件式影像分析方法將光譜資訊及地形參數進行整合後能有效提升崩塌判釋之正確性。;Landslide hazards are common in Taiwan due to its mountainous topography and high number of earthquakes and typhoons experienced yearly. It is essential to develop a method of landslide detection that is capable of providing results with a reasonable level of accuracy, and may also be integrated into an early warning or monitoring system. Object Based Image Analysis (OBIA) is a new method of geographical image investigation which uses segmentation to analyze and process images. Unlike traditional methods, segmentation allows for easy integration of ancillary data during the research process, allowing for the creation of more accurate results. This research attempts to (1) detect and map landslides which occurred following Typhoon Morakot in 2009, (2) incorporate topographic attributes with known effects on the landsliding process (slope, aspect, elevation, curvature, and convexity) directly into the classification process using OBIA and (3) determine the benefits of using segmentation in the landslide mapping process by. A very high resolution 8m FORMOSAT-2 image of the Huaguoshan Basin was used in combination with topographic attributes derived from a 10m digital elevation model for the study area. In order to prove the effectiveness of the topographic attributes in the classification process, the segmentation and classification were first performed with topographic attributes before repeating the process after their removal. The results were subsequently classified based on 6 land use types present within the study area – bare soil, channel, forest, landslide runout and landslide source. Once the classification the accuracy of the process assessed using a random point sampling method. Landslide areas were detected with an overall accuracy of 81.8% and kappa value of 0.64 when using topographic attributes and with an overall accuracy of 77.5% and kappa value of 0.55 without them. The addition of topographic attributes assisted in reducing the amount of misclassifications that occurred in shadowed forest areas and helped separate small urban areas from the surrounding landslide and bare soil areas. This indicates that the integration of topographic attributes is a good means of improving classification accuracy in mountainous areas such as the Huaguoshan basin.
