How Satellite Data Gets From an Instrument to a Usable Product

The gap between what a sensor records and what a customer can use is wider than most people outside the field expect. A satellite does not take photographs in any ordinary sense. It records detector counts, which must be converted into physical measurements, corrected for the instrument's own behaviour, positioned accurately on the Earth's surface, adjusted for the atmosphere in between, and only then combined into something an analyst can interpret. Each of those stages costs time and introduces uncertainty, and the entire chain determines whether the resulting product is worth paying for.

Processing levels, in the order they happen

The convention divides processing into numbered levels, and understanding them clarifies most conversations about data quality. Raw instrument output with telemetry attached is the starting point. Applying calibration converts detector counts into physical units — radiance, brightness temperature, backscatter — while remaining in the geometry of the sensor. Georeferencing places each measurement on the ground, resampling onto a map projection. Higher levels derive quantities of actual interest: surface reflectance corrected for atmospheric effects, vegetation indices, wind vectors, elevation models. Documentation that walks the full chain, such as the product guides at the remote sensing coverage on this site, is worth reading before evaluating any provider's claims, because two products described identically can sit at very different levels.

Calibration is continuous, not a factory step

Instruments change. Detectors degrade under radiation, optics darken, thermal cycling shifts response, and a sensor calibrated perfectly before launch will drift measurably within a year. Maintaining accuracy requires ongoing effort: onboard calibration sources, observations of stable natural targets, cross-comparison with other instruments viewing the same scene. Missions that neglect this produce data that looks fine and cannot support the time-series analysis that most scientific and commercial use depends on. Detecting change over a decade requires confidence that an apparent change is in the world rather than in the instrument.

Georeferencing accuracy limits everything downstream

Placing a pixel correctly requires knowing where the spacecraft was, which way it was pointing, and how the terrain rises beneath it. Errors in any of the three shift the result on the ground. For applications comparing images across dates — monitoring construction, agriculture, deforestation — misregistration between scenes produces false change that no amount of clever analysis can remove. This is why providers quote geolocation accuracy as a headline specification and why users should treat it as one.

• Precise orbit determination, refined after the fact from tracking data
• Attitude knowledge from star trackers and gyroscopes
• A terrain model, whose own errors propagate into position
• Ground control points where available, for the highest accuracy tiers

The atmosphere is between you and everything

Optical sensors measure light that has passed through the atmosphere twice, scattered and absorbed along the way. Correcting for this to recover surface properties requires knowing something about aerosols, water vapour and viewing geometry, and the correction is the largest source of uncertainty in many products. Cloud is a harder problem still: an obscured pixel carries no surface information at all, and cloud masking is an active area of research rather than a solved preprocessing step. Radar instruments avoid both issues by operating at wavelengths that pass through cloud, which is the main reason they remain valuable despite being harder to interpret.

Latency is where commercial value concentrates

For scientific archives, processing time is largely irrelevant. For operational users — disaster response, maritime monitoring, agriculture during a critical window — the time from observation to delivered product is the product. Reducing it means faster downlink, which means more ground stations or relay capacity, and faster processing, which increasingly means performing some analysis onboard so that only conclusions rather than raw frames need to come down. Both are expensive, and providers price accordingly.

Archives, reprocessing and why version numbers matter

Data products are not static. As calibration understanding improves and correction algorithms are refined, providers reprocess historical archives and issue new versions, sometimes changing values enough to alter conclusions drawn from the earlier release. For anyone building an operational system on top of a data feed, this has practical consequences: record which product version each analysis used, expect that a reprocessing campaign will eventually require you to re-run comparisons, and be cautious about mixing versions within a single time series. Providers generally document these changes carefully, and users generally do not read the documentation until a result stops reproducing.

Reading a data specification honestly

Resolution attracts attention and is frequently the least important number. Ask instead about revisit frequency, since a sharp image of the wrong day is useless. Ask about geolocation accuracy and whether it is guaranteed or typical. Ask about radiometric calibration stability, if you intend to compare across time. Ask what the cloud-free acquisition rate looks like for your region, which for some parts of the world is the figure that determines whether the service works at all. And ask about licensing, because restrictions on redistribution and derived products routinely surprise buyers who evaluated only the technical specification.