What we did
RefCal's first research area asks whether effects that shape a measurement show up in real evidence, and whether the conditions under which they were measured are recorded.
Three checks bounded what this set can claim:
- Atmospheric aerosol: a comparison against 87 real AERONET measurements was examined, then excluded as RefCal retrieval evidence when follow-up showed the SIAC output largely reproduced its CAMS prior.
- Viewing geometry: real per-swath viewing and sun angles were extracted from Sentinel-2 product metadata at four reference sites and matched to RefCal's own results at those sites, to check whether a change in viewing geometry between orbits produces the size and sign of effect the model predicts.
- Sensor spectral response: a real, independently measured reference reflectance spectrum was passed through RefCal's Sentinel-2A and Landsat 9 sensor-response operators, both built from the sensors' own official calibration data, to confirm the models connect consistently to a real measurement.
What we saw
The atmospheric comparison did not provide independent evidence that RefCal retrieved aerosol correctly. It measured a third-party processor output that follow-up analysis found was largely locked to its CAMS prior, so this set makes no atmospheric-retrieval claim from those values.
The viewing-geometry check found a real effect with the expected sign at two of the four sites, of a size the geometry-only explanation could not account for on its own; the model showed that one candidate explanation (how the atmospheric processing handles view angle) was too small by more than thirty times to be the cause. Which of the two remaining candidate causes is responsible is not yet settled; see the P02 research area.
The sensor spectral-response check passed: the real reference spectrum produced consistent, physically sensible outputs from both sensor models, confirming that RefCal can connect Sentinel-2 and Landsat readings of the same target on real, measured input.
What it means
The viewing-geometry and sensor-response parts of RefCal's measurement model are grounded in real evidence, not only designed test cases, and their conditions are recorded. The atmospheric comparison shows a boundary instead: a passed-through prior cannot be treated as evidence that RefCal retrieved the atmosphere independently.