Texture analysis and classification of SAR images of urban areas

In SAR image classification texture holds useful information. In a study after the ability of texture to discriminate urban land-cover, a set of measures was investigated. Among them were histogram measures, wavelet energy, fractal dimension, lacunarity and semivariograms. The latter were chosen as an alternative for the well known gray-level co-occurrence family of features. The study was done on the basis of non-parametric separability measures and classification techniques applied to ERS-1 SAR data. The conclusion is that texture improves the classification accuracy. The measures that performed best were mean intensity (actually no texture), variance, weighted-rank fill ratio and semivariogram, but the accuracies vary for different classes. Despite the improvement, the overall classification accuracy indicates that the land-cover information content of ERS-1 leaves to be desired.