Evaluating Super-Resolution Models for Real-World Sentinel-2 Applications: A Case Study
TL;DR
The experiments demonstrate that SR pre-processing improves quantitative fidelity and downstream task performance, enabling low-resolution satellites to compete more effectively with commercial imagery.
Abstract
Abstract. High-resolution Earth observation data are crucial for applications such as agriculture, urban planning, and environmental monitoring. Although commercial satellites provide sub-meter imagery, open-access alternatives like Sentinel-2 are limited to resolutions around 10 m ground sampling distance, which is insufficient for many tasks. In this work, we investigate image super-resolution as a method to bridge this gap, enhancing downstream performance on freely available satellite data. We leverage two 16-bit single-band datasets, consisting of Sentinel-2 (20 m→10 m) and VENμS (10 m→5 m) images, to train and benchmark state-of-the-art SR methods, including transformer- and diffusion-based approaches, across multiple dataset mixes. These models are evaluated quantitatively using reference-based metrics (PSNR, SSIM) using ground-truth and no-reference scores (FID, NIQE) for native upscaling from 20 m→10 m and 10 m→5 m. We observe that different SR architectures present trade-offs between standard quantitative metrics and perceptual image quality. We further assess their impact on a practical downstream task: field boundary detection from Sentinel-2 imagery. Our experiments demonstrate that SR pre-processing improves quantitative fidelity and downstream task performance, enabling low-resolution satellites to compete more effectively with commercial imagery.