Cross-view guided remote sensing super-resolution via constrained residual refinement

Existing RSISR reconstructs high-resolution imagery primarily from the degraded satellite observation, leaving missing fine-scale information underconstrained. We study whether geo-associated but geometrically misaligned ground imagery can provide useful auxiliary information for satellite super-resolution. We formulate Cross-View Guided Super-Resolution (CVG-SR) and propose Selective Residual Constrained Guidance (SRCG), which restricts ground-derived guidance to gated residual refinement around a satellite-conditioned reconstruction anchor. On VIGOR, matched ground guidance improves PSNR by 1.79 dB over the reconstruction anchor under realistic degradation and also outperforms fine-tuned satellite-only baselines. The advantage persists across higher magnification, different reconstruction anchors, and alternative guidance representations, while external-domain evaluations show that the effect remains observable under domain shift. The source code is available at https://github.com/zqluo-buaa/CVG-SR .