A study of the invariance properties of textural features in SAR images

Compares the performance of a number of different texture features for SAR image analysis. These features are derived from the gray-level co-occurrence matrix, local statistics, and lognormal random field models. The features are compared based on: (i) their invariance or robustness with respect to natural changes in SAR signatures; and (ii) their discrimination ability and classification accuracy. The invariance of the texture features is investigated on a set of 13 ERS-1 SAR images of the same scene, captured under different conditions. Two main conclusions can be drawn from this study: (i) the texture features are not invariant with respect to natural changes in the mean backscatter values; and (ii) texture fusion and selection by combining texture features obtained by different models significantly improve the classification accuracy