<p>Polarimetric synthetic aperture radar (PolSAR) image classification remains challenging due to speckle noise, high dimensionality, and limited labeled data. Most existing methods rely on deep learning models, such as convolutional or graph-based neural networks, which require large annotated datasets and significant computational resources. To address these limitations, this work proposes a non-neural graph-based framework that jointly exploits polarimetric, spatial, and structural information. The image is represented as a superpixel-level graph, where discriminative projections are learned using a superpixel-based discriminant analysis embedded into the graph propagation process. In addition, complementary local texture features are extracted using random patch-based representations. The final feature representation is classified using a support vector machine without any deep network training. Extensive experiments on multiple benchmark PolSAR datasets demonstrate that the proposed method achieves superior performance compared to several state-of-the-art approaches. In particular, it obtains overall accuracies of 99.68% on Flevoland and 95.41% on San Francisco, while using significantly fewer training samples than deep learning-based methods. The results confirm that combining graph-based structural modeling with local patch descriptors provides an effective and label-efficient solution for PolSAR image classification.</p>