How Thermal Imaging From Space Actually Works

Sensors, wavelengths, and capturing heat signatures

A thermal satellite image looks simple: a map where bright means hot and dark means cold. Behind it sits a chain that starts with atoms vibrating on the ground and ends with a temperature value in a GeoTIFF. Every link in that chain shapes what you can and cannot see, and most of them trace back to one number.

Everything glows

Anything above absolute zero emits radiation, because the charged particles inside it are constantly moving. The hotter the object, the more it emits, and the shorter the wavelength it emits at.

Two classical laws pin this down. The Stefan–Boltzmann law says radiated energy scales with the fourth power of temperature, so hot things stand out sharply: a surface at 47 °C radiates about 50% more energy than one at 17 °C. Wien’s displacement law says where that energy peaks: around 10 µm for a typical Earth surface, and 0.5 µm for the Sun.

That second number is the one everything else follows from. The Sun peaks in visible light, which is why optical satellites need daylight. The Earth peaks in the long-wave infrared, and it emits there whether the Sun is up or not. Thermal sensors don’t need illumination; they read the planet’s own glow.

The catch: emissivity

There’s a complication buried in the physics. Those laws describe a perfect emitter, known as a blackbody. Real surfaces fall short by a factor called emissivity, which runs from 0 to 1 and varies with material, moisture, roughness and viewing angle.

A satellite doesn’t measure temperature. It measures radiance, the energy arriving at the sensor, and you can only convert that into a temperature if you know the surface’s emissivity. Get it wrong by 0.01 and you’re off by roughly half a kelvin. Get it badly wrong, as you easily can over sand or metal, and you’re off by several.

This is the single biggest source of uncertainty in thermal products, and it explains one of the most common misreadings of a thermal image: a shiny metal roof that looks far cooler than it actually is.

Why the bands sit where they do

“Infrared” spans a wide range, and only part of it is useful for temperature.

The near and short-wave infrared are still mostly reflected sunlight: good for vegetation and minerals, useless for heat. The mid-wave band (3–5 µm) is a transition zone, and it excels at very hot targets like fires and gas flares. The long-wave band (8–14 µm) is the real thermal infrared, sitting on the peak of Earth’s emission, and it’s where most surface temperature work happens.

But the atmosphere doesn’t let all of it through. Water vapour, CO₂ and ozone absorb infrared in broad bands, leaving only gaps, the atmospheric windows, that a satellite can look through. The long-wave window is squeezed by ozone at one end and CO₂ at the other, which leaves a usable stretch of about 10.5 to 12.5 µm. That’s why operational thermal bands sit exactly there, and why they stop at 12.5 rather than running to 14.

And why there are usually two of them. The atmosphere between the ground and the satellite is itself warm and emitting, and water vapour weakens the signal on its way up. Because water vapour behaves slightly differently at 11 µm than at 12 µm, the difference between two adjacent channels tells you how much interference to subtract. That’s the split-window technique, and it’s why Landsat, MODIS, SLSTR and AVHRR all fly thermal bands in pairs.

From radiance to a real temperature

Invert the physics on the measured radiance and you get the temperature a blackbody would need in order to glow that brightly. That’s the brightness temperature, a real physical quantity, but not the temperature of the ground. It still includes the atmosphere, and it still assumes a perfect emitter.

Two corrections close the gap. Atmospheric correction removes the air’s contribution, usually via the split-window difference. Emissivity correction accounts for the surface, using land cover, vegetation indices, or multi-band algorithms. Only then do you have land surface temperature (LST), and the difference between the two can be several kelvin.

Good sensors over well-behaved surfaces land within about 1–2 K of the truth, measured as root-mean-square error (RMSE). Accuracy falls off over mixed terrain, at steep viewing angles, and in humid air.

Clouds are the hard limit. Thermal infrared doesn’t penetrate them, so a cloudy pixel measures the cloud top, which is cold. Miss one in the mask and you’ve recorded a dramatic cold anomaly that doesn’t exist.

Cooled or uncooled

Thermal optics have one unusual constraint: ordinary glass is opaque at 10 µm. Lenses have to be germanium or similar exotic materials, or the design goes all-mirror.

Detectors split into two families, and the choice defines the mission.

Cooled photon detectors turn infrared photons directly into electrical charge. They’re exquisitely sensitive, but the photon energies involved are so small that the detector’s own heat would drown the signal at room temperature, so they need a cryocooler holding the focal plane at a few tens of kelvin. Landsat’s runs at around 43 K.

Uncooled microbolometers work by letting the infrared warm a tiny suspended membrane and reading the change in its electrical resistance. Far less sensitive, but no cryocooler: smaller, lighter, cheaper, and the reason smallsat thermal constellations exist at all.

Neither is better. A climate mission tracking sea surface temperature over decades needs the cooled detector’s precision. A service delivering daily heat maps of a city gets far more from flying twenty cheap satellites than one exquisite one.

Both drift with their own temperature, so thermal instruments regularly look at an onboard reference of known temperature and at deep space, giving two anchor points for calibration on every scan.

What the orbit decides

Thermal pixels are always coarser than optical ones from the same satellite. Longer wavelengths diffract more, so at 10 µm the resolution is roughly twenty times worse than at 0.5 µm for the same telescope. Matching a 30 cm optical pixel in thermal would take an aperture no spacecraft could carry.

After that it’s a three-way trade between pixel size, swath width and how often you come back:

Landsat buys resolution with revisit. Sentinel-3 does the opposite. ECOSTRESS is the interesting one: flying on the Space Station means it isn’t sun-synchronous, so it catches the same place at different hours.

That matters more than it sounds. Most Earth observation satellites cross the equator at the same local time on every pass, usually mid-morning. That’s great for optical consistency, but it means you only ever see one point on the daily heating curve. Peak afternoon heat, evening cooling, pre-dawn minima: all invisible. Thermal constellations increasingly break the convention for exactly this reason.

Reading the picture

A thermal image isn’t a photograph. Each pixel is an estimate, rendered through a colour ramp someone chose, and changing the ramp can make the same data look alarming or unremarkable. A few things surprise people the first time:

Water looks flat and cold, because it changes temperature slowly. At night it often looks warmer than the land around it. That inversion is normal.

Roads and industrial roofs blaze. Dark surfaces that heat quickly light up the image. No other sensor shows the urban heat island this clearly.

Vegetation runs cool, until it stops transpiring. A water-stressed crop warms up days before it looks any different in visible light, which is the whole basis of thermal irrigation monitoring.

Tiny hot things dominate. A gas flare a few metres across can define a 375 m pixel, because radiance at mid-wave scales so steeply with temperature. That’s why fire products use those bands rather than the long-wave ones.

Every pixel is an average. A 100 m pixel over a suburb blends roofs, roads, gardens and trees into one number, which is why resolution matters so much for city and farm applications.

It starts at 10 µm

Everything cascades from one number: the Earth glows at around 10 µm. That sets the band. The atmosphere narrows it and forces you to carry two channels instead of one. Diffraction sets your pixel size. Detector physics sets your price and precision. The orbit sets what you see and how often.

The same chain sets the limits. Thermal works in the dark, often better than by day. It cannot see through cloud, and only radar and microwave sensors can. It describes surfaces rather than objects, and it never sees past a roof or a wall.

Which is why the chain is worth knowing. A thermal image is a measurement, not a picture, and every step between the ground and the pixel leaves a mark. When an analyst finds an oddly cool industrial roof, the first question is whether they’re looking at temperature or emissivity. Both produce the same pixel. Only one says something true about the building.