An exploration of features for SAR classification

Summary form only given. Classification of SAR images usually is constrained by the choice of features to use. Often the powers, in dB, for several frequencies and polarizations are used. However, there are very many other features one could use, some that include the spatial variability between neighboring pixels. This paper explores the utility of using some of these non-standard features in a standard Bayesian classification framework. Some features make the frequency relationships explicit by forming gradients of the standard powers in the frequency domain. Others use the differences between the polarizations and the same or different frequencies. Still others form statistics based on some small neighborhood surrounding each pixel. One such is called texture which is the standard-deviation without the contribution from speckle. Others use the geometrical relationship between the pixels, examples are: lacunarity, grey-level co-occurrence matrices, others?... . For full-polarimetric data there is also available the co-pol phase-difference statistics, which are known to correlate well with the scattering mechanisms associated with the target. There are also less well-known features: (1) polarization for maximum power, (2) polarization for minimum power, (3) power image calculated based on maximum separability between 2 classes, (4) Eigenvalues and vectors of the scattering matrix. For interferometric data there are also some other features: (1) Coherence, (2) Statistics of height information, (3) Differences in height information between different polarizations and frequencies. Each of these features will be evaluated for its usefulness to classification and various optimum scenarios will be presented for different sensor combinations.