SAR images interpretation using data analysis techniques

The paper presents the analysis of the full polarimetric SAR image of San Francisco Bay by the three tools described. In polarimetric data the average Mueller Stokes matrix offers a complete description of backscattering polarisation properties of areas (for fixed frequencies and aspect angles). A particular graphical representation known as the polarisation signature is useful. We describe here a new graphical representation of the information of the Mueller Stokes matrix and a new feature vector. This feature vector is built by analysis of the Mueller Stokes matrix on three basis vectors of Stokes representation space, and by separation of the polarised and unpolarised (fluctuation) part for each backscattered vector corresponding to each basis vector. The projections of these vectors on a polarisation map leads to a new graphical representation where it is very easy to compare different classes (different scattering areas) with their fluctuations. The natural classes are found by a clustering process combining a k-means algorithm and an ascendant hierarchical clustering. The pixels of the image are labelled. A small part of the pixels is used by a learning process. When learned the decision function is applied to all the pixels of the full image. The procedure is mostly based on regression trees techniques and kernel estimates. It supports the following functions: feature automatic selection, Bayesian risk evaluation, extrapolation control, new classes detection, and new classes automatic learning. A generalised classification rate is presented.