PolSAR Image Super-Resolution with Polarimetric Degradation Modulation Network

Most existing deep learning-based polarimetric synthetic aperture radar (PolSAR) image super-resolution methods are trained on paired low- and high-resolution datasets generated by a fixed degradation process (e.g., bicubic downsampling) and do not fully consider the influence among polarimetric channels. Thus, it is challenge for them to reconstruct high-resolution PolSAR images with various degradation forms. To tackle these issues, this work proposed a polarimetric degradation modulation network (Pol-DMnet) for PolSAR image super-resolution. Firstly, a degradation model which takes into account the impacts of blurring, downsampling and speckle effects is developed to represent the resolution degradation process of PolSAR images. Secondly, a polarimetric degradation modulation block is specifically designed to integrate the prior information about PolSAR degradation into the super-resolution process. It enables the network to adapt to various degradation processes. Finally, the Pol-DMnet is developed, by training on paired dataset generated by the proposed PolSAR image degradation model, it is capable of simultaneously integrating the polarimetric and spatial information for PolSAR image super-resolution reconstruction. Comparison studies in terms of various spaceborne and airborne PolSAR datasets are carried out. Compared with the state-of-the-art method, the proposed approach show superiority in both visual and qualitative evaluation. The real resolution enhancements have been inspected by strong points including a deployed corner reflector in terms of the 3 dB width for the first time. The overall resolution has been improved by 1.452 and 1.525 times in range and azimuth directions by the proposed approach which are clearly higher than the comparison methods. In addition, a real measured Radarsat-2 low- and high-resolution PolSAR image pair over the same scene have been used to further validate the practical performance of the proposed Pol-DMnet.