A new POLSAR speckle filtering approach is proposed to overcome the disadvantages of SMBF on polarimetric TSX data. Instead of Freeman-Durden decomposition, the new approach applies dichotomy decompositions (Barnes-Holm) to label each pixel by dominant scattering mechanism as single point target scattering or distributed target scattering. Then the unsupervised Wishart classification is applied to determine the adaptive neighborhood. Finally, only the “N-target” component is involved in the filtering step, while the pure single target component is preserved. By experiment on San Francisco multi-look TSX data, it is illustrated that the new method performs better than SMBF in pixel dominant scattering mechanism labeling and preservation of point target polarimetric properties.