Classification of urban areas from GeoEye-1 imagery through texture features based on Histograms of Equivalent Patterns

Abstract

A family of 26 non-parametric texture descriptors based on Histograms of Equivalent Patterns (HEP) has been tested, many of them for the first time in remote sensing applications, to improve urban classification through object-based image analysis of GeoEye-1 imagery. These HEP descriptors have been compared to the widely known texture measures derived from the gray-level co-occurrence matrix (GLCM). All the five finally selected HEP descriptors (Local Binary Patterns, Improved Local Binary Patterns, Binary Gradient Contours and two different combinations of Completed Local Binary Patterns) performed faster in terms of execution time and yielded significantly better accuracy figures than GLCM features. Moreover, the HEP texture descriptors provided additional information to the basic spectral features from the GeoEye-1's bands (R, G, B, NIR, PAN) significantly improving overall accuracy values by around 3%. Conversely, and in statistic terms, strategies involving GLCM texture derivatives did not improve the classification accuracy achieved from only the spectral information. Lastly, both approaches (HEP and GLCM) showed similar behavior with regard to the training set size applied.