Thermal and Land Surface Temperature

Reflective remote sensing measures a property. Sunlight arrives, the surface returns some fraction of it, and that fraction is a reasonably stable characteristic of the material: a healthy leaf sends back roughly the same share of near-infrared this week as next. Spectral Bands sets out the regions where that logic holds and marks the thermal-infrared as standing outside them.

Thermal remote sensing measures a state. No external illumination is involved. The surface is itself the source, radiating because it has a temperature, and the instrument collects a share of what it radiates. What comes back is therefore not a characteristic of the material but a condition the material happens to be in — one that can shift by ten degrees between mid-morning and mid-afternoon. Almost every difficulty in reading thermal imagery follows from that one substitution of state for property.

What the instrument measures, and what it does not

A thermal band records radiance, and the path from raw sensor output to a calibrated physical value is the same chain Digital Imagery describes for any other band. For a thermal band, though, the chain has somewhere further to go. The relationship between an object’s temperature and the radiation it emits is known for an ideal radiator, so the measured radiance can be worked backwards into a temperature: the temperature an ideal radiator would need in order to emit what the instrument saw. That figure is the brightness temperature, and it is where many thermal products stop.

Brightness temperature is a genuine physical quantity, but it is not the temperature of the ground. It is the answer to a question with two assumptions folded into it — that the surface radiates ideally, and that nothing happened to the radiation on its way up. Both are false, and land surface temperature is what remains after each has been argued past. That makes it a retrieved quantity: derived under stated assumptions rather than read off an instrument. The distinction is not pedantry. It determines what the number can carry.

Emissivity, and why nothing radiates ideally

Every real material emits less than the ideal at a given temperature, and emissivity is the fraction it does emit. Water, snow, and dense vegetation come close to ideal. Dry bare soil, sand, and built materials such as concrete and sheet metal fall further below it and vary far more widely. A surface that radiates less than ideal, read as though it did not, comes out cooler than it really is — and the error is neither small nor random. A few hundredths of emissivity is worth about a degree, always in the same direction, and it lands hardest on the dry, sparsely vegetated, built-up surfaces that thermal questions are most often asked about.

The awkward part is that emissivity is rarely known independently. It changes with material, with wavelength, with how wet a surface is and how it is arranged, so every retrieval has to supply it from somewhere: typical values attached to a classification, an estimate driven by vegetation cover, or a multi-band approach that solves for emissivity and temperature together. Whichever route is taken, an emissivity assumption is inside the number. It is the largest single reason two products over the same place can disagree.

The atmosphere in between

The air is not a clear window even in the region thermal bands occupy. Water vapour is the main absorber, so the size of the correction tracks humidity: modest in cold dry air, substantial over a warm humid landscape. It works in two directions at once. Some of what the surface emits is absorbed before reaching the sensor, which pulls the apparent temperature down; meanwhile the atmosphere radiates in its own right, and part of that downward radiation bounces off the surface into the instrument, which pushes it back up. Undoing both requires an estimate of how much water vapour sat in the column at that moment, and how warm it was.

Cloud is a different matter, and a more dangerous one. Thick cloud is opaque at these wavelengths rather than merely troublesome, so a pixel beneath it reports the cloud’s own temperature, not a weakened version of the ground’s; thin cirrus is subtler still, being partly transparent, and pulls the value down without replacing it. Both are plausible-looking numbers belonging to the wrong thing, and no check on the value alone will catch either. Screening with Cloud Masking is a precondition for a thermal record, not a refinement of one.

Surface temperature is not air temperature

Land surface temperature is the temperature of the radiating skin presented to the instrument: the top of a canopy, a roof, a road surface, the film of a lake. Air temperature, the number in a forecast, is measured in shade a metre or two above the ground. They are related, and they routinely differ by a great deal. Dry sunlit pavement can run far above the air around it; a well-watered, transpiring canopy can sit below it. The gap widens with strong sun, dry soil, and still air.

Which surface the sensor sees is a real question too. Over forest, the value belongs mostly to the canopy top and says little about the ground beneath. Over a city, one pixel mixes roofs, whatever streets are in view, and walls glimpsed at an angle, weighted by geometry rather than by area. A retrieved temperature checked against a weather station will disagree with it, and should.

Coarse pixels and a fixed hour

Two limits follow from the physics rather than from engineering taste. The first is scale: thermal bands are almost always coarser than the reflective bands flying beside them, because the emitted signal available here is far weaker per unit of ground, and the instrument makes that up by collecting over a larger patch. Resolution covers what that trade costs in general; the thermal-specific sting is that temperature varies over very short distances — a shaded courtyard and the sunlit street beyond it can differ by more than the seasons do — so a coarse thermal pixel averages across a spread that the reflective bands never have to span.

The second is timing. Because temperature is a state, the hour of acquisition is part of the measurement in a way it never is for a reflectance. A sun-synchronous platform crosses at close to the same local time on every pass, so an archive such as Landsat samples the same moment of the daily cycle again and again. That consistency is exactly what makes comparison between dates possible — but the record stays silent about every other hour, including the daily peak when it falls elsewhere. Time Series develops spacing and gaps; the point here is the fixed phase.

Reading a retrieved temperature honestly

Because the quantity is retrieved, what a thermal product owes its reader is disclosure. Four things are worth establishing before a value is used: what emissivity was assumed and where it came from, how the atmosphere was handled, which surface the number belongs to, and when it was taken. Two products can differ over the same place and date while both being faithful implementations of different stated choices, so a disagreement between them is not by itself evidence that one is wrong.

This is also why relative use is the safer footing. Comparing one part of a scene with another, or a place against its own history, cancels much of the shared assumption; asserting an exact absolute value at a point rests the whole weight on assumptions the reader cannot see. The Climate page lists this quantity among the things it declines to improvise. What this page adds is the conceptual half of that gap — what a thermal measurement is, and what it costs to trust — and deliberately not the methodological half.