Feature-Enhanced Low-Rank and Sparse Decomposition Network for SAR RFI Suppression

With the increasing number of electromagnetic devices, radio frequency interference (RFI) suppression has gradually become an essential problem in synthetic aperture radar (SAR) imaging. Faced with complex RFI environments such as time-varying and multitype mixing that may occur, traditional approaches often result in inadequate suppression and loss of valuable echoes. Moreover, the representation ability of manually extracted features is limited, struggling to maintain consistent performance in complex electromagnetic environments. To tackle these challenges, this article proposes a feature-enhanced low-rank and sparse decomposition network (FELS-Net), which separates RFI and useful echoes in the time-frequency domain (TFD). We introduce two learnable invertible nonlinear transforms to enhance the representation of RFI and SAR echoes, and unfold the RFI suppression scheme based on low-rank and sparse decomposition into a parameter-learnable network structure. The strong interpretability of model-driven architecture offers a potential stability guarantee for RFI suppression performance, while deep learning (DL) contributes to more effective feature characterization and more efficient and robust parameter schemes. Experimental results demonstrate that the proposed method outperforms comparative approaches in both time-frequency (TF) and image domains, exhibiting robust performance across diverse experimental conditions.