Super-resolution as Spatial Downscaling: Coherence, Structure and Task Utility in Earth Observation

arXiv:2610.23972v1 Announce Type: new Abstract: Learned satellite super-resolution is commonly assessed by pixel fidelity and potential downlink savings, but these criteria do not establish whether the reconstructed field preserves spatial structure or supports reliable downstream analysis. We treat single-image super-resolution as a spatial downscaling problem and evaluate four lightweight networks, including our compact 0.36 million-parameter model, against interpolation using pixel metrics, directional variograms, spectral band correlation, codispersion and observation coherence. On 735 satellite tiles from 190 scenes, the networks differ by at most 0.03dB and outperform bicubic interpolation by 1.39-1.42dB at imes 2. They recover about 30% of the spectral power above the observation's Nyquist frequency and, at imes 4, produce more isotropic fields than the reference (range anisotropy 0.59 against 0.49). Energy-based metrics can favour invented detail. Enforcing coherence reaches the quantisation limit, yet at least 99.7% of the remaining error lies in the observation operator's null space, which coherence cannot constrain. Under spatially blocked, paired evaluation with two backbones and a sharpening control, learned reconstruction changes land-use accuracy by -1.0 to +3.1 percentage points (significant in 7 of 24 comparisons) and building-change F1 by -0.015 to +0.024 (none significant) relative to bicubic interpolation, and amplifies spurious change between two observations of an unchanged scene by 1.3-1.5 times where interpolation attenuates it. Full-resolution JPEG,2000 matches or exceeds reduced-resolution transmission with reconstruction at every matched bitrate. Learned reconstruction is most useful when only low-resolution imagery is available, and its evaluation should combine pixel metrics with spatial structure, observation coherence and task-level tests under spatial validation.

Related

• Semantic-Guided Cross-Sensor Super Resolution of Remote Sensing Images: A Gated Dual Conditioning Flow Matching Model
• Beyond Visual Fidelity: Benchmarking Super-Resolution Models for Large-Scale Remote Sensing Imagery via Downstream Task Integration
• TOC-SR: Task-Optimal Compact diffusion for Image Super Resolution
• GB-LSR: Local Spectral Decoding with a Learned Global Bandwidth for Arbitrary-Scale Super-Resolution

Source: arXiv cs.CV | 2026-10-20