FrugalSR

What it is

Freely available Sentinel-2 optical imagery is capped at 10 m ground sampling — too coarse for fine field, parcel, and land-feature work. Naive super-resolution hallucinates detail that can silently corrupt any measurement downstream. FrugalSR upsamples 10 m RGBN imagery ×4 while enforcing an exact pre-clamp area/nearest projection, so the downsampled super-resolved output reconstructs the original pixels to floating-point precision (measured low-frequency residual ≈ 7.3×10⁻⁹ before clipping over 681 test tiles).

It ships four selectable tiers that trade detail against hallucination risk — each carrying a surfaced risk badge — plus uncertainty, abstention, and input-gate tooling for review. Generated high-frequency detail is treated as an explicit, audited hypothesis, not measured truth.

Why it's different

The open-science frontier (ESA's OpenSR / SEN2SR) has independently converged on FrugalSR's two core ideas — a low-frequency consistency constraint and an explicit hallucination metric — which validates the approach. But it ships as a research library. Commercial upscalers (DigiFarm 10 m→1 m, gamma.earth, LuxCarta) market their output as directly usable with no hallucination audit. FrugalSR occupies the intersection none of them do: a measurement-consistency guarantee + hallucination-audited selectable tiers + abstention, for teams that have to defend their numbers.