Pseudo-Reference-Based Framework for Remote Sensing Image Super-Resolution

Abstract

High-resolution (HR) remote sensing (RS) imagery is essential for accurate scene interpretation but is often difficult to acquire due to sensor limitations and high acquisition costs. While single image super-resolution (SISR) relies solely on low-resolution (LR) inputs and inherently struggles to recover high-frequency details and fine textures, reference-based image super-resolution (RefSR) relies on external HR reference images that are often unavailable or suffer from misalignment in practical RS scenarios. To address these limitations, this letter proposes a pseudo-reference (PRef)-based satellite image super-resolution (SR) framework that eliminates the dependency on external reference images. In Stage 1, a PRef generator (PRG) synthesizes a spatially aligned PRef image from a single LR input. In Stage 2, the PRef-guided SR (PRGSR) network reconstructs the HR image by integrating LR and PRef features through feature extraction, distillation fusion attention (DFA), and reference-aided feature aggregation (RaFA). In particular, DFA complementarily fuses the structural information from the LR and the texture cues from the PRef while suppressing unreliable components. Experiments on RRSSRD and KaggleSRD datasets demonstrate that the proposed method outperforms representative and recent SISR methods in terms of peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) while producing clearer structural details. The results indicate that the generated PRef provides effective and spatially consistent auxiliary guidance, enabling competitive reconstruction performance without requiring external reference images. The source code is available at “https://github.com/WinStone1002/Pseudo-Reference-based-Framework-for-Remote-Sensing-Image-Super-Resolution.git”