傳統的影像分類技術是以像元式(Pixel-Based)的方法為主,但若使用高空間解析度的影像時會產生椒鹽效應(Salt and Pepper Effect)的現象,進而影響分類的精度。基於此,本研究利用物件式(Object-Based)的影像分類方法於Z/I DMC(Digital Mapping Camera)航照影像上萃取出崩塌地區域。首先,本研究比較了Mean-shift和Ward二種不同演算法之影像分割技術,並採取分割結果較佳的Mean-shift演算法將航照影像依同質性分割成不同區塊,接著以每個區塊為最小單元進行分類,最後評估萃取出來之崩塌地的精度。研究結果顯示物件式的影像分類方法可以有效消除在像元式影像分類上所產生的椒鹽效應現象,其崩塌地萃取整體精度從87.05%(像元式分類)提升至99.41%(物件式分類)。
Traditional image classification techniques use per-pixel (pixel-based) approaches to classify images. However, due to "salt-and-pepper effect", these approaches often result in less satisfactory outcome when applied to high resolution aerial image data. Therefore, the objective of this study was to use object-based classification method to detect landslide areas using aerial images acquired by Z/I DMC (Digital Mapping Camera). Firstly, this study compared two kinds of image segmentation techniques (Mean-shift and Ward approaches), and the algorithm with better segmentation result was adopted to segment image into regions based on homogeneity. Then each region was taken as a unit for image classification, and the accuracy of landslide detection was evaluated. The results show that, compared with pixel-based image classification approach, object-based image classification approach can effectively reduce "salt-and-pepper effect" and improve the accuracy of landslide extraction from high resolution aerial images. In this study, the overall accuracy for landslide extraction using pixel-based and object-based classification method was 87.05% and 99.41, respectively.