Polarimetry has been studied for many years in SAR. Due to the enormous quantity of SAR images acquired by satellites or airborne systems, there is an evident need for efficient automatic analysis tools. Classification algorithms are one of the main applications for PoLSAR data. Nowadays, fully polarimetric high resolution sensors can commonly reach up to decimeter resolutions. This yields a higher heterogeneity in the clutter, especially in urban areas, where the clutter can no longer be modeled as a Gaussian process. Recent advances in the field of SIRV (Spherically Invariant Random Vectors) allow the modeling of non-Gaussian clutter as a compound Gaussian process. In this paper, we propose to apply a region growing process as an initialization to a SIRV based classification technique. As the region growing process is shape constrained, spatial features are better delineated and the samples used for the estimation of the coherency matrices are more adapted. Then a statistical clustering technique adapted to the SIRV model is applied to retrieve similarities between regions in the whole image.