Why Do Thermal Cameras Misread Temperature? Three Pitfalls You Need to Know

The 82.4 °C on a thermal camera's screen looks like a measurement. In fact it is a calculation: the camera measures radiant energy, and the temperature is worked back from a few parameters you entered in the menu. Enter the parameters wrong and the number is wrong, while the image looks perfectly normal, with no warning at all.

This article covers three common sources of error, and for the first two works out with real numbers how large the error is.

1. A thermal camera does not actually measure temperature

The energy received by each pixel of the camera comes from three parts:

Wreceived = ε · W(Tobject) + (1 − ε) · W(Tbackground) + atmospheric term

The first term is the radiation the object emits itself, discounted by the emissivity ε; the second is radiation from the surroundings reflected by the object, with coefficient 1 − ε (the reflectance of an opaque object); the third is the air between camera and object, negligible at close range.

To work back to temperature, the camera must first know ε and Tbackground. These two numbers are not measured; you enter them. Enter them wrong and the result is wrong.

2. Pitfall one: the wrong emissivity

Emissivity is "the ability of this surface to emit radiation, as a proportion of an ideal blackbody". It depends on the material, but even more on the surface condition: for the same piece of aluminum, polished and oxidized surfaces differ in emissivity by more than ten times.

Source: FLIR, "How Does Emissivity Affect Thermal Imaging?". The article also points out that surfaces with emissivity below 0.5 basically cannot give accurate temperature readings without surface treatment.

2.1 How far off does a wrong setting get you?

A worked example with the broadband approximation W ∝ T⁴. True object temperature 80 °C, ambient background 25 °C, true surface emissivity 0.85; here is what the camera reads at different ε settings:

Note that the lower ε is, the more the error is amplified. Setting ε to 0.95 in every case:

• Painted surface with an actual ε = 0.90 → reads 77.7 °C, an error of −2.3 °C, acceptable in most cases
• Rough surface with an actual ε = 0.85 → reads 75.3 °C, an error of −4.7 °C, starting to be a problem
• Semi-bright metal with an actual ε = 0.30 (true 200 °C) → reads 109 °C, an error of −91 °C
• Polished metal with an actual ε = 0.10 → reads 32 °C, an error of −48 °C

All of the above use the broadband approximation W ∝ T⁴. A real 8–14 µm longwave thermal camera integrates over its band, so the values differ (for example, for the 200 °C, ε = 0.30 row, band integration gives a reading of about 99 °C), but the direction and order of magnitude of the error are the same.

Measure a metal part with emissivity 0.3 at the default 0.95, and 200 °C is read as 109 °C, while the image looks perfectly normal.

3. Pitfall two: the reflected background changes

The second term of the formula, (1 − ε) · W(Tbackground), says one thing: the lower the emissivity of a surface, the larger the reflected component. A low-emissivity metal surface with ε = 0.15 (for example a lightly oxidized aluminum part) has a reflectance of 0.85; for an object at 80 °C in 25 °C surroundings, about three quarters of the energy the camera receives is not emitted by the object itself but reflected in from the surroundings.

Suppose the camera's background temperature setting stays fixed at 25 °C, the object is at a true 80 °C with ε = 0.15, and the actual surroundings change:

The surroundings changed by only 15 °C, and the reading jumped by forty or fifty degrees. This explains several strange things often seen on site:

• The same spot measures very differently in the morning and afternoon: the factory temperature has changed
• The reading changes when you measure from another angle: what is being reflected has changed
• Odd areas appear "colder than their surroundings": that is cold radiation reflected from the ceiling or windows
• Where the operator stands affects the reading: a person is a 36 °C high-emissivity radiation source

4. Pitfall three: the object does not fill the pixels

The first two pitfalls are radiometric problems; this one is geometric. If the object covers less than a few pixels in the image, each pixel receives a mix of radiation from the object and the background, and the reading is pulled toward the background.

The rule of thumb is: the object must cover at least 3 × 3 pixels before a temperature close to the true value can be measured. To measure a 5 mm component, first confirm how large each pixel is at that distance; this is the IFOV calculation, with complete formulas and examples in the band reference table.

Another number in the datasheet is NETD (noise equivalent temperature difference), the temperature difference that the noise itself corresponds to (the temperature difference at a signal-to-noise ratio of 1); to resolve reliably, a temperature difference usually needs to be 3 to 5 times the NETD. Typical values for uncooled microbolometers are in the tens of mK. Note that:

• NETD is resolving power, not accuracy. For a camera with an NETD of 30 mK, the noise is equivalent to about 0.03 °C and reliably distinguishable differences are on the order of 0.1 °C, but its absolute accuracy is usually still specified as ±2 °C or ±2 %
• NETD is specified under particular conditions (commonly a 30 °C scene with an f/1.0 lens). Change the lens or the scene temperature, and real performance differs

5. How to do it right in practice

The three pitfalls look alarming, but production lines have well-established countermeasures.

5.1 The most effective move: switch to relative judgment

Absolute temperature is hard; relative difference is easy. When parts from the same batch are measured at the same position and angle, emissivity and reflected background are common systematic errors, most of which cancel out on subtraction. So thermographic inspection on production lines often sets its criteria like this:

• ❌ "Alarm if this bearing exceeds 68 °C"
• ✅ "Alarm if this bearing is 12 °C hotter than others in the batch"
• ✅ "Alarm if the temperature difference between three symmetrical positions on this board exceeds 5 °C"

The benefit is that emissivity only has to be consistent within the batch, not precisely known, and the effect of ambient temperature drift is greatly reduced. Note, though, that a wrong emissivity setting still scales the measured temperature difference down roughly in proportion: measure a metal part with an actual ε = 0.30 at 0.95, and a true 12 °C difference reads as only about 5 °C, so thresholds should be set from actual readings under the same settings. Much predictive maintenance and process monitoring needs only relative differences in the first place.

5.2 When absolute temperature is a must

1. Apply high-emissivity tape or matte black paint: stick a small piece of electrical tape of known emissivity on the measuring point, wait for the tape and surface to reach thermal equilibrium, then measure the tape. This is one of the cheapest methods, at the cost of needing access to the object; electrical tape is mostly rated only to about 80–105 °C, so use high-temperature matte black paint for hotter objects
2. Place a reference body: put an object of known temperature and known emissivity in the field of view, and use it to work back to the current environmental parameters
3. Measure the emissivity: heat the object to a known temperature (confirmed with a contact thermometer), and adjust the camera's ε until the reading matches; that ε is the actual value
4. Control the viewing angle: the emissivity of non-metallic surfaces drops noticeably beyond a viewing angle of about 45°–60°, while metals change in a different direction; measure as perpendicular to the surface as possible
5. Avoid reflection sources: keep the camera, lights, windows and operator out of the specular reflection path of the measured surface

Closing: numbers mean something only with their assumptions

A thermal camera is a very useful tool: contactless, full-field and real-time. But every temperature it gives carries a string of assumptions, "assuming emissivity is X, assuming the background is Y". Without those assumptions stated clearly, the number is just a guess that looks very precise.

The good news is that once you know where these three pitfalls are, most measurements can be designed to be unaffected: use relative judgment, apply reference patches, fix the angle and distance. Designing the measurement method correctly is far more effective than buying a more expensive camera.