Deep-Learning and SIRV-Based Dual-Domain Speckle Suppression for PolSAR Imagery

Due to the coherent imaging mechanism of synthetic aperture radar (SAR), speckle noise hinders the interpretation of SAR images. Deep learning networks developed for SAR despeckling exhibit promising noise reduction capabilities, but it is difficult to apply the deep-learning-based model on the polarimetric SAR (PolSAR) data directly. Polarimetric information is often distorted by networks due to the complex-valued channels or the normalization applied to PolSAR data. In this paper, we propose a deep learning despeckling network for PolSAR data based on the spherically invariant random vector (SIRV) model which decomposes the polarimetric coherency matrix into a normalized polarimetric coherency matrix and a texture component. Two network models are developed for polarimetric domain and texture domain separately. A block combining Swin-Transformer and convolution is used to extract global and local features of the texture, and a Complex Block was used in polarimetric domain for feature extraction of complex inputs. A new loss function of ratio balanced mean square error (RBMSE) is proposed for the texture domain filtering to help the network handle unnormalized data. PolSAR data in different frequencies and resolutions are used to illustrate the robustness of the proposed method. The results show that the proposed deep-learning and SIRV-based dual-domain filtering algorithm achieves good performance on images acquired by different sensors. It provides a new framework for deep learning based filtering of PolSAR image.