A Joint Spatial–Frequency Network with a Multi-Dimensional Loss Function for Remote Sensing Image Super-Resolution

Remote sensing image super-resolution (SR) aims to recover high-resolution images from low-resolution observations, but existing deep networks often emphasize spatial-domain modeling and pixel-wise optimization, which may limit their ability to restore high-frequency textures and preserve perceptual structures. To address these limitations, this study proposes a joint spatial–frequency domain SR reconstruction network (JSF-SRNet) and a joint loss function with pixel-level, multi-scale, and global constraints (JLF-PMG). JSF-SRNet uses a dual-branch design: the spatial branch extracts local multi-scale contextual features, while the frequency branch models global contextual information to complement spatial representations. JLF-PMG supervises reconstruction from three perspectives, including pixel fidelity, multi-scale structural consistency, and global image consistency. The method was trained and evaluated on the public AID dataset and compared with classical and state-of-the-art SR networks. At 2×, the proposed method achieved 38.48 dB PSNR, 0.936 SSIM, and 0.039 SAM, with a 0.22 dB PSNR gain over the second-best method. At 4×, it achieved 33.35 dB PSNR, 0.819 SSIM, and the best NIQE of 6.34, with a 0.04 dB PSNR gain while using only 0.393M parameters. These results demonstrate the effectiveness of integrating spatial- and frequency-domain representations with multi-dimensional loss constraints under the controlled AID benchmark setting.