A technique is presented for multiscale texture analysis and segmentation of polarimetric SAR images. Textural features are extracted using a multiscale wavelet decomposition based on a wavelet frame. The feature vector is composed of local variance estimates of the smooth image and of the wavelet coefficients. The decomposition is performed at two scales and using images derived by polarimetric power synthesis at a set of polarization configurations. This set is chosen based on a priori-knowledge of the texturally optimal polarization states. Alternatively a complete and nonredundant representation of the full polarimetric information consisting of nine backscatter intensities is used. Feature reduction is achieved by an approximate solution of the Multiple Discriminant Analysis (MDA) transform. A set of controlled experiments, based on Monte Carlo simulations, is set up to assess the performance of the technique with respect to texture segmentation problems. One case is reported concerning the simulation of a fragmented forest, where two vegetation classes with different structural characteristics are mixed. Finally, as an example of the application of the technique to real SAR data, texture segmentation of a high resolution image acquired by the DLR E-SAR sensor at L-band is illustrated.