Research Of High Resolution Remote Sensing Image Classification Based On Texture

The extraction of texture is one of the key steps in high resolution remote sensing image classification. Extracting robust and discriminating texture is important to improve the image classification accuracy. According to characteristics of high resolution remote sensing image, we propose a new classification method combine with texton method in this paper, and we also study the extraction of local and global textural feature respectively. The main research work of this paper can be summarized as follows:1) According to the various, changeable and complicated texture in high resolution remote sensing images, a method for extracting the local texture based on random projection is proposed. First, rotation invariant texture vector is obtained by sorting the raw image patch; then it is projected into a compressed space by random projection, which can reduce the feature dimension and keep the information.2) According to the importance of spatial information to the global texture description, a method for representing global texture based on sorted visual words co-occurrence matrix (VWCM) is proposed. First, we reduce the dimension of VWCM by the second clustering; then, we describe the relative spatial information by VWCM by statistic the occurrence of words in different direction; and we sort the VWCM according to the self co-occurrence counts to emphasis the "key words" and improve discriminating of texture and improve classification accuracy.3) The high resolution remote sensing image is classified by integrating local and global texture and combining with support vector machine. In the experiments, we analyze the optimum scale problems for local and global texture, and the affection of the dimension of features to classification accuracy. Our method is proved to be effective and better than traditional texture extraction methods by comparative analysis.