Object-Oriented Land Cover Image Classification System

Yun      Yongsheng; Chang      Qingrui; Xie      Jing

doi:10.2174/187221210790244721

Abstract

In order to develop a new object-oriented image classification method with fuzzy support vector machines for land cover, an effective fuzzy membership as a function of fuzzy nearness is used for reducing the effect of outliers in sample sets. Firstly, according to the spatial and spectral characteristics of different targets on rectified image, the number of objects was automatically determined by using mean shift algorithm, in which local objects were picked up with arbitrary shapes and unique mode labeling. Then, a comparison to other object-oriented methods, which were standard support vector machines (SVM) and K nearest neighbor (KNN), without such pre-processing was successively validated. Finally, the comparison was also made between the traditional pixel-based algorithm and the proposed approach. A high precision object-oriented recognition system is established for remote sensing images. Experimental results indicate the proposed method is much more accurate than those traditional pixel-based algorithms and object-oriented algorithms without pre-processing in the study region.

Keywords: Land cover, object-oriented image classification, mean shift segmentation, fuzzy nearness