Polarimetric Synthetic Aperture Radar (PolSAR) provides rich scattering information that is highly valuable for land-cover classification. However, two major challenges remain: redundancy among polarimetric features and the presence of salt-and-pepper noise in classification maps. In this study, we propose a novel PolSAR classification framework that integrates an iterative correlation-based feature selection strategy with a rank-based post-processing approach. The iterative method progressively eliminates the most redundant features while retaining complementary and informative descriptors, thus preserving a larger and more discriminative feature space compared to conventional one-shot elimination. To evaluate its effectiveness, four machine learning algorithms—K-nearest neighbours (KNN), support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGB)—were applied to a Gaofen-3 PolSAR dataset acquired over San Francisco. The results show that the proposed feature selection approach consistently improves classification performance, with accuracy gains of up to 4% across classifiers. Furthermore, applying a median filter as a post-processing step significantly enhances spatial coherence, achieving accuracies as high as 0.99 for the XGB classifier. These findings confirm that the proposed framework effectively addresses both feature redundancy and spatial noise, leading to more reliable and robust PolSAR classification outcomes.