Physics-Guided Polarimetric Feature Selection Based on Dual-Branch Networks for PolSAR Agriculture Field Classification

Deep learning methods have been widely applied to Polarimetric Synthetic Aperture Radar (PolSAR) land cover classification. However, their performance is often constrained by information loss and the presence of visually similar patterns caused by similar scattering mechanisms. In contrast, directly leveraging rich polarimetric decomposition features introduces challenges due to their high redundancy and the difficulty of identifying the most discriminative feature subsets effectively. To address these issues, we propose a novel Physics-guided Dual- Branch Network. This network comprises a spatial branch that processes Pauli images and a polarimetric branch dedicated to processing 23-dimensional polarimetric features. Within the polarimetric branch, we design a Local Group-Adaptive Feature Selection mechanism that dynamically selects discriminative features during end-to-end training. Crucially, we introduce a Multi-Family Diversity Constraint that groups polarimetric features according to their underlying physical scattering mechanisms. We enforce Group Regularization, Sparsity Regularization, and Physical Correlation Regularization to ensure that the selected features are physically consistent and complementary. Experiments conducted on two agricultural datasets (AIRSAR and RADARSAT-2) demonstrate that the proposed method significantly outperforms state-of-the-art approaches in terms of Overall Accuracy (OA) and Mean Intersection over Union (mIoU). Furthermore, visualization analysis confirms that the selected features closely align with the scattering mechanisms of different crop types, substantially enhancing both classification performance and physical interpretability.