Earth observation

Earth observation is the largest scientific and commercial use of satellites after communications. This page works from the physics of what can be measured through to the data products that result, because the constraints at the top propagate all the way down.

The diffraction limit

Optical resolution is bounded by physics, not by engineering effort. A circular aperture of diameter $D$ observing at wavelength $\lambda$ cannot resolve angular detail finer than

At orbital altitude $h$, that becomes a ground sample distance of

The third row matches a real system closely: WorldView-3 flies a 1.1 m aperture at 617 km and delivers 0.31 m panchromatic imagery. Using 450 nm rather than 550 nm reflects that the panchromatic band spans roughly 450 to 800 nm, and the sharpest detail is carried by the shortest wavelengths.

Three consequences follow, and each explains something about the industry.

Resolution requires aperture. Sub-metre imaging needs a metre-class telescope, which needs a large, stiff, thermally stable structure. This is why high-resolution imaging satellites are large and expensive and why they cannot be built as CubeSats regardless of how good the electronics become.

Flying lower helps linearly. GSD scales with altitude, so halving altitude halves GSD for the same optic. That is the argument for very low Earth orbit, and it is bounded by atmospheric drag, which rises far faster than linearly. See Station keeping.

Regulation, not physics, is often the binding limit. Commercial imaging resolution has been restricted by licensing in several jurisdictions, historically at 0.5 m and later relaxed. See Policy and regulation.

The trilemma

Resolution, swath width and revisit interval cannot be optimised together. The relationship is not a soft trade but close to a hard constraint set by detector count and data rate.

A pushbroom imager has a linear detector array of $N$ pixels perpendicular to the ground track. Swath is then simply

so for a fixed detector array, finer resolution means a narrower swath, proportionally. A narrower swath means fewer square kilometres collected per orbit, which means longer to cover a region, which is worse revisit.

Working the numbers for a WorldView-class system at 0.31 m GSD and 13 km swath:

Ten gigabits per second is the raw output of a single high-resolution optical sensor. Nothing downlinks that continuously, which is why high-resolution satellites image in short tasked bursts rather than continuously, and why onboard compression is mandatory rather than optional.

The resolution of the trilemma in practice is not to resolve it. Operators run two different kinds of system and sell both:

These are complementary rather than competing. A change detected in daily medium-resolution imagery triggers a tasking request for high-resolution imagery, and the two-tier architecture is now the standard commercial model.

Synthetic aperture radar

SAR escapes the aperture limit entirely, which is why it is worth its complexity.

A real-aperture radar has the same diffraction problem as an optic, and at centimetre wavelengths it is far worse: resolving 1 m at 700 km with a 5 cm wavelength would need a real antenna kilometres long.

The synthetic aperture technique instead uses the satellite’s own motion. As the platform flies, it illuminates a target from many successive positions. By recording the phase as well as the amplitude of each return and coherently combining them, the processor synthesises an aperture as long as the distance flown while the target was illuminated.

The result is startling: azimuth resolution equals half the physical antenna length, and is independent of range and wavelength. A 10 m antenna gives 5 m azimuth resolution whether the target is 500 km away or 800 km away.

Range resolution comes from a different mechanism entirely, the bandwidth of the transmitted chirp:

A 300 MHz bandwidth gives 0.5 m slant range resolution. The two resolutions are therefore set by completely independent parameters, which is unlike any optical system.

The final row is what makes SAR scientifically indispensable rather than merely weather-independent. Interferometric SAR compares the phase of two acquisitions of the same scene. Because phase is sensitive to a fraction of a wavelength, differences reveal ground displacement at millimetre scale.

This is how subsidence, volcanic inflation, glacier flow, landslide creep and earthquake deformation fields are measured from orbit. No optical technique approaches that sensitivity, because it depends on coherent phase, which an incoherent optical image does not carry.

SAR’s costs are real. The transmitter needs hundreds of watts to kilowatts, which drives the power and thermal subsystems, and is a large part of why SAR satellites often fly dawn-dusk sun-synchronous orbits for continuous illumination. See Power systems.

Spectral sensing

The distinction between multispectral and hyperspectral is qualitative, not just a band count. Multispectral samples the spectrum at a few chosen points and can classify broad categories. Hyperspectral records a near-continuous spectrum for every pixel, which allows identification by matching against laboratory spectral libraries: a specific mineral, a specific crop, a methane plume.

The cost is photons. Splitting the same incoming light across 200 narrow bands instead of 4 broad ones leaves each band with far less signal, so hyperspectral instruments trade spatial resolution and signal-to-noise for spectral detail. This is a fundamental radiometric trade, and it is why hyperspectral imagers have coarser GSD than multispectral ones on comparable platforms.

Pansharpening exploits the same trade in reverse. A satellite collects a high-resolution panchromatic band and lower-resolution multispectral bands, then fuses them: spatial detail from the panchromatic, colour from the multispectral. Most commercial colour imagery marketed at a given resolution is produced this way.

Processing levels

Raw sensor output is not usable data. The standard hierarchy is worth knowing because it determines who does what work and where the value sits.

Two corrections do most of the work between L1 and L2.

Orthorectification removes the distortion caused by terrain relief. A pixel imaged obliquely over a mountain is displaced from its true ground position by an amount proportional to elevation, and correcting it requires a digital elevation model. Geolocation accuracy is limited by the accuracy of that model and by the satellite’s attitude knowledge, which is the point made in GNC about knowledge mattering more than accuracy.

Atmospheric correction converts top-of-atmosphere radiance into surface reflectance by modelling scattering and absorption along the path. Without it, images of the same scene on different days are not comparable, because the atmosphere changed even if the ground did not. Any change detection worth trusting is done on atmospherically corrected data.

The downlink bottleneck

The defining operational constraint of Earth observation is that satellites generate far more data than they can transmit.

The gap is more than an order of magnitude. Four mitigations are used, usually in combination:

Duty cycling

Image only when tasked over a target. High-resolution satellites are idle most of the time, which is why tasking is the commercial product rather than raw coverage.

Onboard compression

Lossless gives roughly 2:1, visually lossless wavelet compression 4:1 to 10:1. Standardised through CCSDS 122 and 123.

Polar ground stations

A station at high latitude sees a polar-orbiting satellite on nearly every orbit rather than a few times a day, roughly tripling contact time. This is why Svalbard exists as a commercial ground station site.

Onboard processing

Downlink the detection rather than the image. Ship a ship position, a cloud mask, or a change flag instead of the pixels that produced it.

The last of these is the one changing fastest, and it inverts the traditional architecture. Rather than downlinking everything and analysing on the ground, run the analysis onboard and downlink the answer. The economics are compelling when the answer is kilobytes and the imagery is gigabytes. See Edge computing in space and Satellite image processing.

Applications

The commercial market has shifted from selling images to selling answers. The value in counting cars in retail car parks or detecting illegal fishing is not in the pixels but in the derived indicator, delivered as a number on a schedule. This changes what the space segment must optimise: revisit and timeliness matter more than resolution once the resolution is sufficient to detect the object of interest.

References

• Elachi, C. and van Zyl, J. Introduction to the Physics and Techniques of Remote Sensing, 3rd ed., Wiley, 2021.
• Jensen, J. R. Remote Sensing of the Environment: An Earth Resource Perspective, 2nd ed., Pearson, 2006.
• Cumming, I. G. and Wong, F. H. Digital Processing of Synthetic Aperture Radar Data, Artech House, 2005.
• Ferretti, A., Prati, C. and Rocca, F. “Permanent scatterers in SAR interferometry”, IEEE Transactions on Geoscience and Remote Sensing, 39:8, 2001.
• CCSDS 122.0-B and 123.0-B, image data compression recommended standards.
• Updike, T. and Comp, C. Radiometric Use of WorldView-2 Imagery, DigitalGlobe technical note, on calibration and processing levels.