Rotation invariant texture feature algorithms for urban settlement classification

The Human Visual System (HVS) has been observed to process visual information on a multi-channel filtering basis in the early stages of analysis. This has given rise to a number of texture segmentation techniques that seek to mimic the HVS multi-channel filtering theory. In this paper we present one such technique, Gabor Filters (GF). We apply Gabor Filters to the identification of the different textural regions in urban areas. GF have received much attention due to their ability to achieve localization in the spatial and spatial frequency domain. With their ability to be tuned to specific spatial frequencies and orientations, they can be used to mimic the HVS approach to identify the different texture regions. Along with the Gray Level Co-occurrence Matrix (GLCM), we analyze the performance of the traditional GF and GLCM compared to their rotation invariant counterparts. Specifically, GF and GLCM are applied to the analysis of QuickBird imagery of an area in Soweto (Johannesburg, South Africa) to identify and classify the different settlement types found in the area. Then the resulting feature vector is analyzed using the Correlation Feature Selector (CFS) algorithm to reduce the number of features used. The results showed that GF perform better in the classification of urban areas, traditional or rotation invariant when compared to GLCM.