Abstract Super-resolution (SR) of Synthetic Aperture Radar (SAR) imagery is necessary to increase the significance of low-resolution Sentinel-1 data in applications such as urban monitoring, land cover analysis, and flood mapping. However, standard SR models are hampered by the intrinsic noise, low contrast, and structural ambiguity in SAR data. In this paper, we propose SARGANet, a hybrid deep learning architecture that combines the structural-preserving qualities of U-Net with the generative capabilities of Generative Adversarial Networks (GANs). Our model leverages a U-Net-based generator, supplemented with spatial attention and lightweight dense blocks to capture fine-grained textures and edge data, while adversarial training encourages perceptually and structurally accurate outputs. SARGANet combines multi-stream inputs (such as SAR-VV and SAR-VH) in a dual-encoder framework to better leverage cross-polarization information without cross-band interference. The network is trained with a composite loss function that is optimized for SAR features and combines pixel-wise, perceptual, adversarial, and structural similarity terms. Tested on a carefully selected Sentinel-1 dataset, SARGANet demonstrates consistent improvements in PSNR and SSIM compared to several baseline and state-of-the-art super-resolution methods under the evaluated experimental setup. The groundwork for impending multi-modal super-resolution frameworks merging the optical and SAR domains is provided by this study, which also highlights the potential of GAN-U-Net fusion for practical SAR applications.
Talreja et al. (2026) studied this question.