Single-Image Super-Resolution of Sentinel-2 Imagery for Filling MisrSat-2 Acquisition Gaps Using Deep Learning Techniques

Abstract

Satellite imagery is of great importance in Earth observation, particularly in highly populated countries like Egypt, where environmental monitoring and urban planning are essential. The country has recently launched MisrSat-2, a high-resolution satellite capable of providing images with a spatial resolution of 2 m; yet its limited temporal coverage leads to considerable gaps in observation and therefore hampers its usefulness for monitoring rapid changes on the ground. This represents a major technical limitation since, for any given date, all the high-resolution spatial data is lost if there is no image acquired by MisrSat-2. To address this shortcoming, the current study employs deep learning-based super-resolution (SR) techniques to generate 2.5 m proxy images that resemble MisrSat-2, using freely available Sentinel-2 imagery (at 10 m), thereby helping to fill the temporal coverage gaps without the need for additional satellite acquisitions. It must be emphasized that the outputs are model-generated proxies and not real 2.5 m observations. Four different SR architectures were systematically evaluated: RCAN, SRGAN, ESRGAN, and SwinIR. The RCAN model obtained the best performance (PSNR = 26.629 dB, SSIM = 0.750), beating SRGAN (PSNR = 23.35, SSIM = 0.716), ESRGAN (PSNR = 24.38, SSIM = 0.735), and SwinIR (PSNR = 22.68, SSIM = 0.744), although the differences in SSIM values between the models are minor and the order of performance reflects the conditions under which the evaluation was carried out. Since the evaluation is conducted on histogram-matched imagery, the reported values relate to the structural and perceptual quality of the reconstruction rather than to absolute radiometric accuracy. A key aspect of this study is the development of a fully automated co-registration framework between Sentinel-2 and MisrSat-2 images, which does not require any manual input and which corrects spatial misalignments while considerably cutting down the preprocessing time. As MisrSat-2 is a satellite that has only recently been launched, this study is the first to apply and test deep learning super-resolution methods using MisrSat-2 as the reference, and it offers a practical and scalable approach to obtaining higher-resolution Earth observation over Egypt.