Atmospheric correction estimates how a surface would appear without the scattering and absorption between it and a satellite sensor. The detector observes radiance, from which a calibrated top-of-atmosphere quantity may be produced. Bottom-of-atmosphere or surface reflectance is then retrieved through a physical model. It is not a direct measurement of the ground.
From sensor signal to surface estimate
Molecules and aerosols change an optical observation through scattering, while gases including water vapour and ozone absorb radiation. A correction processor combines the satellite observation with viewing and illumination geometry, estimates of atmospheric state, terrain and a radiative-transfer model. Sentinel-2’s Sen2Cor processor derives aerosol optical thickness and water-vapour maps before retrieving surface reflectance from look-up tables.1
Measured radiance couples the atmospheric and surface contributions, so the inversion cannot completely separate them from that value alone. Sen2Cor therefore makes modelling choices, including Lambertian surface reflectance and a constant viewing angle within each 30 × 30 km sub-scene.1 Actual surfaces can reflect differently by direction, while aerosol and water vapour vary across a scene. The result is a model-conditioned estimate whose assumptions and processing baseline belong in its provenance.
Copernicus Sentinel-2 Level-2A packages orthorectified surface reflectance with scene classification, cloud and cloud-shadow classes, aerosol optical thickness and water-vapour layers.2 Those supporting layers are part of using the reflectance correctly rather than optional decoration.
Uncertainty and quality screening
Correction quality changes with atmosphere, surface type and geometry. Low Sun angles lengthen the atmospheric path; bright, snow-covered, coastal and cloudy scenes can make aerosol or surface retrieval difficult. The Sen2Cor algorithm document warns that observations above a 70° solar-zenith angle are processed using a clipped angle, which can under-correct the atmospheric signal.1
Validation must be tied to a processor version. A 2018 Sentinel-2 quality report, based on Sen2Cor 2.4 and 2.5.3, found that 97% of evaluated water-vapour retrievals met their stated requirement but only 39% of aerosol-optical-thickness retrievals did.3 This is evidence about those versions and tests, not a performance claim for every later baseline.
Quality flags remain necessary after correction. Landsat surface-reflectance products likewise include per-pixel quality bands for adverse instrument, atmospheric and surface conditions, including cloud contamination.4 Analysts should retain these masks, scale factors and processor metadata, and should treat invalid or high-uncertainty pixels as missing rather than as corrected measurements.
References
Footnotes
-
Copernicus Sentinel programme, Sentinel-2 Level-2A Algorithm Theoretical Basis Document. ↩ ↩2 ↩3
-
Copernicus Sentinel programme, Sentinel-2 Collection 1 MSI Level-2A. ↩
-
Copernicus Sentinel programme, Sentinel-2 Level-2A Data Quality Report. ↩
-
US Geological Survey, Landsat Surface Reflectance Quality Assessment. ↩