A deep learning approach to jointly super-resolve and despeckle Sentinel-1 imagery

Synthetic Aperture Radar (SAR) systems enable all-weather, day-and-night imaging. Sentinel-1 (S1), a widely used SAR constellation, provides free imagery but is constrained by its limited spatial resolution and level of speckle noise, which affects interpretability. While commercial SAR satellites offer higher-resolution imagery with reduced speckle noise, their prohibitive costs restrict accessibility for many remote sensing applications. Although various SAR despeckling and super-resolution methods exist, mainly based on deep learning, they typically focus on one of these task rather than addressing both simultaneously. Moreover, previous SAR super-resolution methods often rely on generative models, which may introduce hallucinated (invented) artifacts and learn the noise distribution of the reference images, ultimately preserving speckle noise, rather than suppressing it. Additionally, these methods typically use high-resolution commercial data as ground truth (e.g. TerraSAR-X), potentially leading to discrepancies in intrinsic sensor characteristics. To overcome these limitations, we propose using non-generative neural networks — commonly used for single-image super-resolution — to jointly super-resolve and despeckle S1 Interferometric Wide Swath mode imagery, leveraging S1 Stripmap mode as a reference. By avoiding generative mechanisms, our approach prevents the model from learning and propagating noise patterns, effectively suppressing speckle while enhancing spatial resolution from 10 m to 5 m Ground Sample Distance. Additionally, using the same sensor in a different operational mode, ensures greater dataset consistency, eliminating discrepancies in sensor characteristics. Experimental results demonstrate that our method effectively doubles the resolution of S1 imagery while reducing speckle noise. • Proposed a method for simultaneous resolution enhancement and speckle noise reduction in S1 IW imagery. • Ensured dataset consistency by utilizing the S1 data in different operational modes. • Applied non-generative deep learning single-image super-resolution architectures. • Evaluated alternative approaches to preserve the input data distribution during training.