| The topic is interesting to us as we (especially my co-founder) have done lots of research in this area, both for adapting existing and inventing new Super Resolution methods for satellite images. We kinda discussed this in detail yesterday. Btw, we have a demo of the result of the enhancement we achieved about 2 years ago at https://demo.valinor.earth Looking at this implementation, the noise artifacts likely stem from a few sources. The RCAN model normalizes input by dividing by 400, which is a fixed assumption about Sentinel-2’s radiometric range that doesn’t account for atmospheric variability or scene-specific characteristics. Plus, working with L1B data means you’re enhancing atmospheric artifacts along with ground features - those hazy patterns aren’t just sensor noise but actual atmospheric scattering that gets amplified during super-resolution. Over the past 2 years, we’ve hit several walls that might sound familiar: - Models trained on clean datasets (DIV2K, etc.) completely fall apart on real satellite imagery with clouds, shadows, and atmospheric effects. - The classic CNN architectures like RCAN struggle with global context - they’ll sharpen a building edge but miss that it’s part of a larger urban pattern. - Training on one sensor and deploying on another is impossible without significant degradation. Some fixes we’ve found effective: - Incorporate atmospheric correction directly into the SR pipeline (check out the MuS2 benchmark paper from 2023). - Use physics-informed neural networks that understand radiative transfer. - Multi-temporal stacking before SR dramatically reduces noise while preserving real features. For anyone diving deep into this space, check out: - SRRepViT (2024) - achieves similar quality to heavyweight models with only 0.25M parameters. - DiffusionSat - the new foundation model that conditions on geolocation metadata. - The L1BSR approach from CVPR 2023 that exploits Sentinel-2’s detector overlap for self-supervised training. - FocalSR (2025) with Fourier-transform attention - game changer for preserving spectral signatures. Also worth exploring is the WorldStrat dataset for training, and if you’re feeling adventurous, the new SGDM models claiming a 32x enhancement (though take that with a grain of salt for operational use). The real breakthrough will likely come from models that jointly optimize for visual quality AND radiometric accuracy. Current models excel at one or the other, but rarely both. If you interested in these topics, we would love to connect. We are at brajeshwar@valinor.earth and amir@valinor.earth |