A Quad-Polarimetric SAR image reconstruction network and wetland classification fusing optical texture and polarization information

Wetland mapping is crucial for ecological security, yet its application is frequently constrained by the limited availability of Quad-Polarimetric (QP) Synthetic Aperture Radar (SAR) data. In contrast, widely accessible Dual-Polarimetric (DP) SAR datasets are characterized by information deficiency, making the differentiation of complex vegetation types an ill-posed inverse problem. To address this research gap, this study proposes a novel network named QPSAR-RecNet. Different from conventional purely data-driven approaches, QPSAR-RecNet leverages high-resolution PlanetScope (PS) optical textures as structural priors to regularize the inversion space. Specifically, the Texture-Guided Attention (TGA) module integrates optical spatial cues into SAR feature learning, while a Physics-Aware Loss function guarantees the Positive Semi-Definiteness (PSD) of reconstructed covariance matrices and thereby ensures the physical interpretability of model outputs. Experiments carried out in the source area of the Eastern Route of the South-to-North Water Diversion Project (SNWDP) demonstrate that QPSAR-RecNet can effectively retrieves key scattering components. Quantitative assessments reveal that the QPSAR-RecNet-based classification achieves an Overall Accuracy (OA) of 91.15%, which is markedly higher than the Sentinel-1 (S1) baseline model (82.68%) and comparable to the RADARSAT-2 (RD-2) QP reference standard (92.45%). Moreover, sensitivity analyses validate the framework’s robustness against both speckle and thermal noise. By combining the physical properties of SAR data with the structural details of optical images, QPSAR-RecNet provides an economical and efficient solution for large-scale ecological monitoring.