A Joint Real- and Complex-Valued Network for Classification of Pol(In)SAR Images

Synthetic Aperture Radar (SAR) systems capture both amplitude and phase information, producing complex-valued images widely used in Earth observation applications. Among SAR modalities, Polarimetric SAR (PolSAR) and Polarimetric Interferometric SAR (PolInSAR) systems leverage polarimetric and interferometric information, typically represented by a data coherency matrix comprising real-valued diagonal elements and complex-valued off-diagonal elements. Existing deep learning approaches process the entire coherency matrix either through purely real-valued or purely complex-valued networks, which fails to fully exploit its heterogeneous structure. In this paper, we propose a structurally decoupled modeling strategy for coherency matrices, which explicitly separates and processes diagonal and off-diagonal components based on their distinct structural roles. A real-valued network is employed to model the diagonal scattering power components, while a complex-valued network is used to capture the cross-channel correlation, orientation, and coherence structures from the off-diagonal components. This design aligns well with the inherent organization of the SAR coherency matrix, enabling more targeted and effective feature learning. The extracted real and complex features are subsequently fused via a cross-domain enhancement fusion block to achieve robust representation learning. Experiments on DLR's FSAR dataset demonstrate that the proposed method consistently outperforms six state-of-the-art techniques across three different SAR modalities, achieving superior performance in both quantitative accuracy and qualitative robustness.