Analysis and Comparison of Texture Feature Extracting methods for High Resolution RS images

摘要

Texture is one of the important features on remote sensing (RS) images, especially, for high resolution RS image in which plays an important role of remote sensing image classification. The texture feature has proven to be of great help to the information extraction by researchers in previous literatures. In order to evaluate a suitable method to a certain scale of images and feature patterns we made a comparison and analysis of various texture feature extracting methods. This paper paid more attention to the signal processing approaches like LBP, Gabor and Log-gabor as processing signals in frequency domain is much more matched to the mechanism of human visual perception. But gray level co-occurrence matrix (GLCM) and variogram approaches are also implemented during the comparison since they are respectively used to the nature objects and artificial objects extractions in many applications. Further we proposed an approach for extracting and supporting use of texture information within a Shape-Adaptive Neighborhood (SAN) to improve the classification accuracy of the regional applications with high resolution RS image. However, all the traditional texture feature extraction methods can not fit the SAN as they followed regular window (rectangle or square) calculations since SANs are usually given irregular objects. Based on the characteristics of this novel SAN algorithm, this paper further describes the procedure of texture extraction, and advances the concrete implementation methods of gray level co-occurrence matrix (GLCM), variogram, LBP and Log-gabor. The former two give a solution of the irregular shape problem by modifying and comparing them for SAN. Experimental results shown, it needs to modify data in the co-occurrence matrix directly with coefficient S but not necessary to simulate the data based on the statistic characteristics of GLCM and variogram. Here, the coefficient S is defined by the quotient of neighborhood size and vision window size. And the SAN modeling procedure can just modify the windows size of LBP and log-gabor. Finally, we compared the results of modified algorithms of LBP, Log-gabor, GLCM and variogram. The applied experiments show that modified extraction method under the SAN (shape-adaptive neighborhood) of remote sensing image feature extraction method makes the results of RS image classification accuracy increased around 4% (GLCM) to 8% (LBP).