This paper proposes a new object-based classification method for Polarimetric synthetic aperture radar(Pol SAR) images, which considers scattering powers from an improved model-based polarimetric decomposition approach, as well as the spatial and textural features. With the decomposition, the scattering ambiguities between oriented buildings and vegetation are reduced. Furthermore,various contextual features are extracted from the object and incorporated into the K-nearest neighbors(k-NN)based classification. To reduce the feature redundancy, a new Supervised locally linear embedding(S-LLE) dimensionality reduction method is introduced to map the high dimensional polarimetric signatures into the most compact low-dimensional structure for classification. Experimental results with Airborne synthetic aperture rada(AIRSAR)C-band Pol SAR image demonstrate the superior performance to other methods. This paper proposes a new object-based classification method for Polarimetric synthetic aperture radar (Pol SAR) images, which consider scattering powers from an improved model-based polarimetric decomposition approach, as well as the spatial and textural features. With the decomposition, the scattering ambiguities between oriented buildings and vegetation are reduced. Further, various contextual features are extracted from the object and incorporated into the K-nearest neighbors (k-NN) based classification. To reduce the feature redundancy, a new Supervised locally linear embedding (S- LLE) dimensionality reduction method is introduced to map the high dimensional polarimetric signatures into the most compact low-dimensional structure for classification. Experimental results with Airborne synthetic aperture rada (AIRSAR) C-band Pol SAR image demonstrates the superior performance to other methods.