Physically consistent multitemporal dual-domain learning for dual-polarization SAR despeckling

Speckle contamination strongly affects the interpretability and practical application of synthetic aperture radar (SAR) imagery. For real dual-polarization SAR data, despeckling is particularly difficult because clean supervision is unavailable and temporal variations between acquisitions complicate the characterization of speckle behavior. Existing unsupervised deep models have shown promise, but many of them still underuse the complementary information contained in neighboring temporal observations and insufficiently exploit the physical cues associated with polarimetric scattering, which may result in structural degradation and distortion of scattering characteristics. To address these limitations, we develop CD 2 P-Net, a collaborative dual-domain progressive network for real dual-polarization SAR despeckling. The method is established within a physically guided multitemporal unsupervised learning framework, where spatial, spectral, and temporal relationships are progressively modeled to make better use of temporal coherence and speckle variability across adjacent acquisitions. In addition, covariance-derived statistics and polarimetric decomposition descriptors are introduced to improve the physical reliability of the learned features. Experiments on real Sentinel-1 dual-polarization time-series data show that the proposed method achieves superior despeckling performance compared with representative competing approaches, while better preserving spatial structures, polarimetric characteristics, and semantic consistency.