Learned satellite super-resolution is usually evaluated by how closely reconstructed pixels match the original and by how much data could be saved in downlink. According to a new arXiv paper, these criteria do not establish whether the reconstructed field actually preserves spatial structure or supports reliable downstream use. The paper argues that such metrics can be misleading when the goal is to produce physically coherent imagery.
The authors suggest reframing super-resolution as a form of spatial downscaling, where the focus shifts to structural coherence and task utility. Instead of asking only whether pixels look correct, evaluation should ask whether the reconstructed field maintains the spatial patterns needed for Earth-observation applications. This aligns with a broader trend in remote sensing toward application-aware model assessment.
The abstract is truncated, so the specific methods and experimental results are not available from the source. Still, the core argument is clear: the field needs evaluation criteria that go beyond pixel-level accuracy and data savings, and that directly address whether super-resolved products are fit for purpose.