Deep learning method enhances Sentinel-1 radar images, beats MERLIN baseline
New self-supervised framework improves SAR image quality without needing ground truth data.
Synthetic Aperture Radar (SAR) imagery is critical for all-weather, day-and-night Earth observation, but suffers from speckle noise and artifacts. Sentinel-1's Stripmap (SM) mode offers the highest resolution among its imaging modes, yet spatial constraints and noise limit finer-detail applications. Existing enhancement methods often rely on external sensors or simulated ground truth, which are not always available.
Now, researchers from multiple institutions present a self-supervised framework that leverages azimuth subaperture decomposition to create paired training data directly from the full-aperture image. The approach integrates single- and multi-frame learning with an iterative inference scheme that progressively refines quality. Experiments on real Sentinel-1 SM data show consistent outperformance of the widely used self-supervised baseline MERLIN in terms of PSNR and SSIM, though MERLIN achieves higher Equivalent Number of Looks (ENL), indicating a trade-off between structural fidelity and speckle smoothing. The framework is reproducible and extendable to other SAR platforms, polarizations, and acquisition modes. The paper was accepted at the AI4Space Workshop at CVPR 2026.
- Self-supervised framework uses azimuth subaperture decomposition to generate training data without external references.
- Outperforms MERLIN baseline in PSNR and SSIM on real Sentinel-1 Stripmap data.
- Method extendable to other SAR platforms and modes; accepted at CVPR 2026 AI4Space Workshop.
Why It Matters
Improves satellite radar image quality without costly ground truth, enabling better Earth monitoring and defense applications.