Investigating Blind Image Super-Resolution of Sentinel-2 Satellite Data and Its Applications

Abstract

High-resolution Earth observation data are crucial for applications such as agriculture, urban planning, and environmental monitoring. Although commercial and expensive satellites can capture sub-meter imagery, open-access alternatives like Sentinel-2 are limited to resolutions around 10m, which is insufficient for many applications. In this thesis, we investigate image super-resolution (SR) as a method to bridge this resolution gap, improving the performance of downstream tasks on freely available satellite data. We developed two 16-bit single-band datasets with different spatial resolutions, using Sentinel-2 (20m → 10m) and VENμS (10m → 5m), with the goal of training and benchmarking four different super-resolution methods. To this end, we adapted three transformer models (SwinIR, Mat, PFT) and one diffusion model (EDiffSR) to our unique satellite data. After training them with three different dataset mixes, we evaluated their performance quantitatively utilizing standard reference-based metrics (PSNR, SSIM). With FID and custom-trained NIQE models, we assessed the native upscaling capabilities of all twelve model configurations. In addition, we evaluated their impact on a practical downstream application, a Sentinel-2 field boundary detection. Our experiments demonstrate that the Transformer models performed well in terms of PSNR and SSIM, as well as in our downstream application, proving the value of using super-resolution as a preprocessing step. EDiffSR achieved sharper and perceptually more realistic imagery, outperforming our Transformers on FID and NIQE, but failed to beat bicubic upsampling on our downstream task. These findings highlight that super-resolution can be used to make low-resolution satellites more competitive against commercial imagery.