Polarimetric Synthetic Aperture Radar (PolSAR) imagery, as an advanced remote sensing technology provides rich information about the scattering characteristics of the Earth's surface, significantly enhancing classification accuracy. However, effectively integrating contextual features with polarimetric ones remains a key challenge. In this paper, we propose a framework for PolSAR image classification that leverages graph modeling to capture spatial relationships and utilizes the scattering characteristics with physical interpretability of polarimetric data. The proposed method begins with superpixel segmentation to reduce computational complexity and maintain spatial homogeneity. Then, superpixels construct the graph nodes. To enhance class separability, we suggest a superpixel based discriminant analysis transformation to compute the weight matrix used in the graph propagation. Unlike deep learning approaches that learn weights through complex neural networks, our method uses the discriminant analysis based projection to derive the weight matrix in a physically interpretable and computationally efficient manner. The graph based projection results in spatial-polarimetric features that encode both structural and discriminative information. In parallel, we extract another set of polarimetric features from the H-A-Alpha (Cloude-Pottier) decomposition describing the physical scattering mechanisms. The final classification is performed by fusing the graph-derived features and the decomposition-based features and feeding them into a standard classifier such as support vector machine.