Abstract
Remote sensing image super-resolution (SR) techniques play a crucial role in geographic information analysis, environmental observation, and urban development planning. However, existing approaches are computationally intensive, which hinders them from bewing applied on resource-constrained devices. Although numerous efforts have focused on efficient image SR, the intrinsic sparsity characteristics of remote sensing images remain under-explored. To tackle these challenges, this paper introduces an efficient SR method founded on a dynamic Sparse Swin Transformer (S2Transformer). First, a dynamic sparse mask module is proposed to distinguish important regions from other ones. Subsequently, a dynamic sparse Transformer is developed to adaptively focus on important regions with more computational resources being allocated, markedly reducing redundant computations over background regions. Experiments are conducted on several benchmark remote sensing datasets and the results demonstrate that the proposed approach significantly outperforms existing methods in detail restoration, edge sharpness, and robustness, achieving superior PSNR and SSIM scores.