Quantitative Analysis of Second Order Statistical Class Feature for VHR Remote Sensed Imagery.

Abstract: The progresses of statistical texture analysis have been playing a more and more important roles in remote sensing mapping, pattern recognition and computer vision. The achievements and research approaches have found important applications in many fields, including industry inspection, remote sensed data analysis and mapping, medical imaging, textile defect detection, video image analysis, food grading, and natural texture recognition and retrieval, object recognition, etc.[1] This paper starts with the principles and algorithms of second order statistics. Based on the theoretical analysis, we selected 5 typical texture class samples from Quick Bird RGB fused data with 0.61m resolution. We used GLCMs to quantitatively calculate texture features, which parameter values are suitable for the specific texture classifications. GLCMs of 5 typical texture class samples from the data set were calculated. Six statistical features for every class sample in four orientations and 1 pixel of pair-wise distance were obtained, including: energy, entropy, contrast, homogeneity, correlation, and dissimilarity respectively. The average values in four directions were computed and compared. The results show that dissimilarity and entropy have biggest value differences among six samples. They are the most important features for classification or recognition of class samples. The statistics of dissimilarity, entropy, homogeneity, contrast have been demonstrated a decrease in classification ability. But the average contrast can discriminate the complex building sample from other textures in spite of small differences with others. The results of the research supplied important references for the quantitative interpretation of VHR Quick Bird imagery in the applications of land cover/use classification and mapping. [Copyright &y& Elsevier]