Temporal Variation-Guided Self-Supervised PolSAR Despeckling Network

Abstract. Deep learning–based polarimetric SAR (PolSAR) despeckling faces two main challenges: the scarcity of clean reference images and the difficulty of preserving structural details while suppressing noise. To address these issues, we propose a temporally guided self-supervised network (TGSD-Net), which generates pseudo training pairs from consecutive noisy observations and leverages a change detection–based prior to exploit temporal redundancy and enhance robustness to land-cover changes. TGSD-Net further integrates model and input feature refinements, including auxiliary polarimetric decomposition parameters and a spatiotemporal information fusion module (STIFM) based on a U-Net backbone, to improve temporal and scattering feature representations. The network is specifically designed to robustly handle multi-temporal SAR acquisitions and heterogeneous land-cover types, maintaining consistent scattering structures across different scenes. Extensive experiments on real PolSAR datasets demonstrate that TGSD-Net effectively balances noise suppression with detail preservation. Quantitative metrics, including the equivalent number of looks (ENL) and edge preservation degree (EPD), confirm its superior despeckling performance. Polarimetric decomposition analyses further verify that the network preserves the physical scattering characteristics of PolSAR images.