Object-oriented land cover mapping in China national geographical conditions census

Abstract

Remote sensing image classification is one of the important methods of obtaining land cover/use information for nationwide general survey of geographical conditions in China. In this study, an object-oriented decision tree algorithm is utilized for land cover and land use mapping from high resolution WorldView-2 satellite remote sensing images. The major steps of this approach include image segmentation to get image objects and pixel-based classification based on the image objects. Multi-resolution segmentation approach was adopted in this study, which utilizes spatial and spectral information of land covers in the image to segment into different small objects with particular structural, spectral and texture attributes. A new decision tree classifier called AdaTree. Weight Leaf (AdaTree. WL) was applied to conduct classification process based on the segmented objects, which is modeled by combining the algorithms C4.5 and AdaBoost. The decision algorithm is integrated in the software--GLC (Global Land Cover Classification) classifier. Several study sites are selected from northwest and southeast in China to test GLC classifier, and compared to SVM (Support Vector Machine), the GLC classifier could reaches better results (mean kappa coefficient is 84.61%) in different scenarios with WorldView-2 images.