A Combined Use of Decomposition and Texture for Terrain Classification of Fully Polarimetric Sar Images

This paper presents two-stage unsupervised terrain classification of fully polarimetric SAR data using Freeman and Durden decomposition based on three simple scattering mechanisms: surface, volume and double bounce (first step), and textural features (uncorrelated uniformity, contrast, inverse moment and entropy) obtained from grey level co-occurrence matrices (GLCM) (second step). Textural features are defined in moving window 5x5 pixels with N=32 (N – number of grey levels). This algorithm preserves the purity of dominant polarimetric scattering properties and defines textural features in each scattering category. It is shown better object discrimination after applying texture within fixed scattering category. Speckle reduction is one of the main moments in image interpretation improvement because of its great influence on texture. Results from unfiltered and Lee filtered polarimetric SAR images show that the values of contrast and entropy decrease and the values of uniformity and inverse moment increase with speckle reduction, that’s true for all polarizations (HH, VV, HV). The discrimination between objects increases after speckle filtering. Polarization influence on texture features is defined by calculating the features in SAR images with HH, VV and HV polarizations before and after speckle filtering, and then creating RGB images. It is shown more polarization influence on texture features (uniformity, inverse moment and entropy) before filtering and less influence – after speckle filtering. It’s not true for contrast where polarization influence is not changed practically with filtering. SIR-C/X-SAR SLC L-band images of Moscow region are used for illustration.