From orbit to ground truth

01A pixel is a measurement

A satellite image looks like a picture, but each pixel is a measurement of radiance or backscatter from a patch of ground, taken at a known time from a known position.

Landsat satellites have acquired images of the land surface continuously since 1972, building an archive that land managers, planners and scientists use to understand change².

Sentinel-2 samples 13 spectral bands at 10, 20 and 60 metres, across a swath 290 kilometres wide³. Sentinel-1 carries a C-band radar that images day and night, regardless of the weather⁴.

Copernicus, the Earth observation component of the European Union space programme, draws on satellite data and on in situ measurements from the ground, the air and the sea⁵.

That pairing is the point. Images from orbit cover the whole planet, but only measurements on the ground can tell whether a pixel says what it seems to say.

Between the sensor and the user sit many steps: the atmosphere the light passed through, the angle of the sun, the geometry of the view and the processing that turns counts into physical units. Each step can add error, and each needs a check.

Calibration and validation are the two disciplines that close that gap. Calibration defines how a sensor responds to known, controlled signals; validation assesses the quality of derived products by independent means¹.

02Calibration: tying the sensor to known signals

Every instrument drifts. Optics darken, detectors age and electronics change with temperature, so a sensor that was right at launch will not stay right without help.

Calibration starts on the ground, against laboratory references, and continues in orbit. Sentinel-2 carries an on-board diffuser that collects sunlight for radiometric calibration of its imager³.

Its published performance states an inter-band calibration accuracy of 3 per cent and a multi-temporal calibration of 1 per cent³.

Vicarious calibration adds a check from outside the satellite. Pseudo-invariant sites, instrumented test sites and intercomparisons between sensors all test the instrument against something it cannot influence⁶.

The quality assurance framework endorsed for Earth observation asks for data with a quality indicator traceable to reference standards, preferably the international system of units⁷.

When one satellite replaces another, the overlap matters. Sentinel-2C operated alongside Sentinel-2A before replacing it, to keep calibration seamless and the record unbroken³.

03Validation: checking products against the ground

Calibration makes the radiance right. Validation asks a different question: is the derived product, such as surface reflectance, soil moisture or burned area, right where it matters?

The answer must come from independent means. The calibration and validation community defines validation as assessing the quality of products by means that do not depend on the system that made them¹.

In practice that means field measurements taken at the same place and close to the same time as the overpass: radiometers on masts, soil probes, plots measured by hand and instruments carried by crews.

The match-up is harder than it sounds. A pixel covers tens of metres while a probe sees a few centimetres, and clouds, shadows and timing all open gaps between the two.

Good validation sites are chosen for that reason. They are uniform over many pixels, measured often, documented carefully and kept running long enough to see seasons and sensor changes.

Validation is never finished. Seasons change the surface, sensors age, and processing improves, so the same site is measured again and again to show whether the product still holds.

Validation results then travel with the product. Users can compare the stated uncertainty with their own needs, which is the idea of fitness for purpose that runs through quality assurance for Earth observation⁷.

04Drones between orbit and ground

A drone flies at a height where its pixels are centimetres across, yet it can cover a whole validation site in one flight. That makes it a bridge between a satellite pixel and a point on the ground.

Landsat 9 images at 30 metres in its visible and shortwave infrared bands, across a scene 185 kilometres wide⁸. A field plot measured by hand may be only a few metres across.

A drone survey timed with the overpass shows how uniform the site really is inside each satellite pixel. It reveals the edges, gaps and mixtures that a few point measurements would miss.

Drone sensors need the same discipline as satellite sensors. They are calibrated against reference panels, flown in stable light, and their own uncertainty is recorded with the data.

Used this way, the stack works as one instrument. Points on the ground anchor the drone map, and the drone map tests the satellite product across its full footprint.

Flights must also respect the rules of the airspace and the privacy of people below. Plan routes, heights and data handling with the same care as the measurement itself, and record them with the survey.

05From downlink to analysis-ready data

Ground truth is only half of the ground segment. The other half receives the data from orbit, processes it and keeps the chain of evidence intact on the way to users.

Open space data standards set how that happens. Synchronisation and channel coding for telemetry links are specified so that space-to-ground links protect the data as it is received⁹.

Image compression is standardised too, so that two-dimensional image data from payload instruments can be compressed on board and decompressed correctly on the ground¹⁰.

After processing, analysis-ready data are satellite data processed to a minimum set of requirements, organised for immediate analysis and interoperable through time and with other datasets¹¹.

Each step adds metadata: the calibration used, the processing version, the quality flags and, for validated products, the uncertainty found against the ground. That metadata is part of the measurement, not a label on it.

Field data needs the same treatment on its way in. Ground instruments and drone surveys should arrive with their own time, position, calibration and quality flags, so the match-up compares like with like.

06From calibrated pixels to maps and models

Analysis-ready data are only the start. Most people meet Earth observation as a map, a layer in a geographic information system or the output of a model. Every one of those is built from rasters: grids of pixels, each tied to a place on the ground.

Rasters travel well when their format carries the georeference with the pixels. GeoTIFF does this inside an ordinary TIFF file, and it is an Open Geospatial Consortium standard¹².

Cloud Optimized GeoTIFF arranges the same file in tiles and reduced-resolution overviews, so a web client can use HTTP range requests to read only the part of the image it needs¹³. A map can open a continent and a field from the same file without downloading either in full.

Finding the right image is its own problem. The SpatioTemporal Asset Catalog specification gives collections of imagery a common description of where, when and what, so they can be indexed and searched in the same way¹⁴.

Machine learning now turns these rasters into maps of land cover, water, crops or damage, a field often called GeoAI. A model learns from labelled examples, and those labels are ground truth too: they need a known place, time and method, like any validation measurement.

A model's map is a derived product, so it needs the same independent validation as any other¹. Accuracy reported against held-out ground observations, with the date and region it applies to, should travel with the map.

07Keeping a record consistent for decades

The value of Earth observation grows with the length of the record. A trend in forest cover, coastline or snow is only as good as the consistency of the measurements behind it.

Missions are therefore planned as series. Landsat 9 was launched into the orbit of Landsat 7 to maintain coverage, and joins Landsat 8 to observe each location every 8 days⁸.

Sentinel-1C was launched in December 2024 and re-established a two-satellite radar constellation once its commissioning ended⁴.

The global climate observing community lists the essential climate variables that observations must sustain, from soil moisture and land cover to fire and above-ground biomass¹⁵.

Every replacement satellite, processing update and new validation site is a chance to break the record or to strengthen it. Overlap, cross-calibration and published validation are how it is strengthened.

A record is also a promise to future users. Someone studying change decades from now will depend on the notes kept today about every instrument, calibration and processing choice behind each pixel.

08Designing for ground truth

Plan validation with the product, not after it. Decide which variables matter, what uncertainty users need, and which sites can show whether the product meets that need.

Choose sites that are uniform, accessible and safe to visit through the year. Instrument them continuously where possible, and visit them during overpasses with drones and crews.

Reference everything to the same position and time. A match-up between a pixel and a probe is only as good as the agreement on where and when each was taken.

Publish the method and the results openly. The calibration and validation community exists to share protocols, facilities and results so that different missions can be combined¹.

Nexus Observatory Inc. designs observing sites and data services that can serve as ground truth for imagery, with every value carrying its time, place and quality.

Keep the sites running after the opening campaign. The long record is what reveals drift, and drift is what calibration and validation exist to catch.