Radar image classification and multiresolution analysis

Unsupervised Bayesian segmentation applied to the whole radar image gives good results only when the look number is sufficiently high to approximate the intensity image gamma probability density by a Gaussian law. This is not the case for real remote sensing radar images as ERS1 where the look number is four. Multiscale image analysis by wavelets has proved to be efficient for increasing classification performances in the Gaussian case or for optical images. It is proposed to extend the method to radar images with low look numbers. In unsupervised Bayesian classification, the knowledge of the class a priori laws and of the texture conditional densities is needed. When a Gaussian density is assumed for the speckled texture, it remains Gaussian through the scale pyramid created by the multiscale analysis. In the radar case, the problem lies in the fact that the gamma law is transformed through the scale filters into other densities, difficult to express. They are approximated by the Gram Charlier development, which corrects the Gaussian density by using cumulants of orders 3 and 4. This method has been applied to simulated radar images. A three level multiresolution analysis has been chosen and a comparison was made between four cases: classification without multiscale analysis, with Gaussian hypotheses, with the Gram Charlier approximation, with a Gaussian hypothesis on the third level and approximation on the second of the scale pyramid