Remote Sensing Image Super-Resolution Using Enriched Spatial-Channel Feature Aggregation Networks

Single Image Super-Resolution (SISR) pertains to the intricate task of recovering high-fidelity, high-resolution images, a process highly germane to remote sensing applications. Conventional Convolutional Neural Network (CNN) paradigms have achieved commendable results through the incorporation of residual architectures and attention mechanisms. However, these approaches are mired by excessive localization and information attrition due to simplistic stacking of residual structures and naive embedding of attention modules. This work introduces a sophisticated Transformer-based architecture, specifically tai-lored for super-resolution in remote sensing imagery, adept at harnessing high-dimensional spatial and channel features. The crux of this methodology is the integration of Local Enhanced Transformer layers and Global Transformer layers, adept at cap-turing multiscale features across spatial domains. Additionally, the Residual Channel Attention Block (RCAB) is employed for channel-wise feature calibration. Central to our approach is the introduction of a Spatial and Channel Attention fusion(SCF) module, ingeniously designed to exploit the synergies between spatial and channel information. Comprehensive experimentation substantiates the efficacy of the proposed SCANet, outstripping the performance of contemporary state-of-the-art alternatives in super-resolution tasks.