PCA-Aware Attention Feature Fusion With Complex-Valued Adaptive Weighted UNet for PolSAR Classification

Polarimetric synthetic aperture radar (PolSAR) possesses a unique capability to capture multi-channel scattering information, demonstrating significant value in land cover classification tasks. To fully exploit both the amplitude and phase information embedded in the PolSAR imagery, this paper proposes a land classification method that integrates the complex-valued modeling with feature enhancement strategies. Specifically, a principal component analysis-based attention-aware feature selection module (PCAAF) is introduced to select and assign weights to the information-rich real-valued polarimetric features. These selected features are then fused with the original complex-valued PolSAR data along the channel dimension through a complex- and real-valued fusion strategy, enhancing the overall semantic representation. Furthermore, an improved complex-valued adaptive weighted U-Net (CV-AW-UNet) is designed to incorporate an adaptive feature weighting (AFW) module to dynamically regulate multi-channel features, thus enhancing the feature discriminability. Experimental results on three representative PolSAR datasets, namely San Francisco, Flevoland, and Oberpfaffenhofen, demonstrate that the proposed method outperforms existing the state-of-the-art models in terms of the average accuracy (AA), overall accuracy (OA), and Kappa coefficient, exhibiting superior robustness and generalization ability. These results validate the significant potential of combining complex-valued modeling with attention-guided feature selection in improving land cover classification performance for PolSAR imageries.