MTSR-GAN: achieving 2.5 m resolution from 10 m Sentinel-2 images with a novel super-resolution GAN framework
Yunhe Li, Mei Yang, Tao Bian, Haitao Wu · International Journal of Remote Sensing · 2025
The Sentinel-2 satellite provides optical remote sensing imagery with a maximum resolution of 10 metres, which is inadequate for many applications. To address this limitation, this study proposes the MTSR-GAN method, aimed at enhancing the resolution of satellite images to 2.5 metres. In response to the scarcity of high-resolution images, we utilize K-GAN to estimate degradation kernels and inject noise, generating realistic low-to-high-resolution image pairs. Furthermore, we design an MTSR architecture that integrates the TDSS module from Mamba and overlapping cross-attention blocks based on the Transformer. Experimental evaluations demonstrate that MTSR-GAN surpasses existing methods in multiple no-reference image quality metrics, providing clearer details and fewer artefacts. Visual comparisons also highlight MTSR-GAN’s ability to generate high-resolution images with excellent perceptual quality, indicating its potential for significant contributions in remote sensing applications.