EO SAT GPT
Home › Technical Guides › Data Engineering
Data Engineering

Radiometric Calibration and Orthorectification for Analysis-Ready Data

Why raw Level-1 satellite pixels cannot be stacked directly into maps or AI pipelines, and how geometric plus radiometric processing unlocks interoperability.

From Raw Telemetry to CEOS Analysis Ready Data

Raw downlinked imagery contains uncalibrated Digital Numbers (DNs) distorted by sensor optics, spacecraft attitude jitter, and terrain relief. Overlaying unorthorectified Level-1 files directly onto maps shifts roads and ports by tens to hundreds of meters.

Analysis Ready Data (ARD) following CEOS-ARD standards delivers orthorectified, calibrated products ready for immediate multi-date stacking and AI inference.

  • Level 1B: Sensor-space imagery with Rational Polynomial Coefficients (RPCs).
  • Level 1C: Terrain-corrected map-projected pixels in a standard CRS.
  • Level 2A (ARD): Orthorectified plus atmospherically corrected surface reflectance.

How Orthorectification Removes Terrain Distortion

Orthorectification projects pixels from slanted focal planes onto true map coordinates using sensor RPCs, a Digital Elevation Model (DEM), and optional Ground Control Points (GCPs).

DEM resolution governs residual error: pairing 30m Copernicus DEM or local LiDAR with automated GCP tie-points pushes circular positional error (CE90) below 2 to 5 meters.

  • Match DEM accuracy to off-nadir angle to bound horizontal relief shift.
  • Distinguish bare-earth DTMs from building-inclusive DSMs ('true ortho').
  • Standardize local metric distance measurements on appropriate UTM zones.

Radiometric Calibration to Physical Reflectance

Raw DNs reflect electronic gain settings rather than physical ground brightness. Gain/bias coefficients and vicarious desert-playa calibrations convert DNs into physical Top-of-Atmosphere (TOA) radiance and reflectance.

Normalizing solar angle and Earth-Sun distance prevents seasonal brightness drift from skewing multi-date AI classifiers.

Resampling and Cloud Optimized GeoTIFF (COG)

Reprojecting grids requires choosing resampling kernels: nearest-neighbor preserves discrete QA/cloud mask integers, while bilinear or cubic convolution smooths continuous imagery.

Packaging outputs as tiled Cloud Optimized GeoTIFFs (COGs) with internal overviews lets cloud workers stream exact byte ranges without downloading full scenes.

Satellite Processing Levels & Readiness

LevelGeometric StateRadiometric StateRecommended Use
Level 0Raw telemetry framesUncalibrated streamGround archive only
Level 1BRPCs attached, unprojectedRelative calibrationCustom stereo photogrammetry
Level 1CDEM orthorectifiedTOA reflectanceVisual basemaps & mapping
Level 2A (ARD)Sub-pixel aligned gridSurface reflectanceMulti-date AI & indices
COG + STACInternal tiles & pyramidsIncludes QA mask bandsCloud-native API streaming

Exact level labels vary slightly across commercial providers; verify geometry and reflectance specs in STAC metadata.

Worked Illustration: Terrain Displacement at 25 Degrees Off-Nadir

A satellite images a hillside depot at 1,200m elevation under a 25-degree off-nadir angle without DEM terrain correction.

  1. Compute Factor: tan(25 deg) = ~0.466.
  2. Unadjusted Error: 1,200m x 0.466 = ~559m horizontal map displacement if flat ground (0m) is assumed.
  3. Apply 30m DEM (~4m vertical accuracy): Residual terrain shift drops to 4m x 0.466 = ~1.86m.
  4. Refine with Tie-Points: Matching road corners pulls sub-pixel drift under 1 pixel.

Worked Outcome: DEM orthorectification shrinks a 559-meter hillside shift down to sub-2-meter alignment.

Orthorectification & ARD Checklist

  • Confirm whether delivered scenes are Level 1B sensor-space or orthorectified.
  • Validate CE90 positional accuracy against authoritative reference road vectors.
  • Verify DEM coverage across coastal ports and high latitudes.
  • Use nearest-neighbor resampling for QA/cloud masks; bilinear/cubic for RGB.
  • Convert outputs to tiled Cloud Optimized GeoTIFFs (COGs).
  • Retain RPC and radiometric metadata inside STAC sidecars.

Frequently Asked Questions

How does orthorectification differ from simple georeferencing?

Georeferencing warps a flat plane to control points, whereas orthorectification models 3D viewing angles and DEM elevation to remove terrain relief shift.

Do tall buildings still lean after standard orthorectification?

Yes, under bare-earth DTMs building roofs lean outward unless a 3D Digital Surface Model ('true ortho') is used.

Why use Cloud Optimized GeoTIFF (COG)?

Internal tiling and overviews let web viewers fetch only needed map areas and zoom levels via HTTP range requests.

Can NDVI be calculated from raw Digital Numbers?

No; convert DNs to TOA or preferably BOA surface reflectance first so sensor gains and haze do not distort band ratios.

Primary Standards & References

  • CEOS — Analysis Ready Data (CEOS-ARD) Product Family Specifications
  • OGC — Cloud Optimized GeoTIFF (COG) Standard 1.0
  • ISO 19130-1 — Geographic Information Imagery Sensor Models for Geopositioning

Related Earth Observation Guides