Thermal-infrared processing is challenging
Why are thermal-infrared (TIR) images difficult to process?
TIR geometric challenges
You can expect unique geolocating challenges when processing TIR images.
Lack of reference data
TIR imagery often lacks reliable reference data for feature matching. Reference data is used to align images with a known accurate image, collecting tiepoints and geolocating an image precisely.
Using reference data with similar spectral bands is better for matching. Sentinel-2 and Landsat-8/9 are commonly used as reference sets for VNIR data. Let’s look at how useful these are for TIR matching.
Sentinel-2 SWIR
The Sentinel-2 short-wave infrared (SWIR) band offers a 20-meter resolution. It only correlates well with TIR data under certain conditions. This is because SWIR (1.3–3 µm) and TIR (8–14 µm) pick up different physical signals. In urban areas, SWIR and TIR data correlate quite well, because the contrast pattern between concrete rooftops and tar streets is similar in both.
Landsat-8/9
Landsat-8/9 has its own TIR bands. It is a better spectral match for typical TIR bands (example, 8– 14 µm). However, it is limited to a coarse 100-meter resolution. This caps the achievable geolocation accuracy at roughly that same level.
TIR imagery is coarse
TIR sensors typically produce large pixel sizes, often reaching resolutions several times coarser than visible-to-near-infrared (VNIR) imagery. This makes it difficult to find distinct features to match. A road crossing, for example, may not be visible at all in an image with 100 m resolution. A problem that was small for processing VNIR imagery becomes a big problem for processing TIR imagery.
The resolution difference between TIR imagery and VNIR imagery (about 9x coarser).
TIR reference imagery is time-sensitive
TIR imagery is extremely time-sensitive, since surface temperatures can change rapidly within hours. Features in a TIR image are distinguished by their temperature outline and no longer by their visible shape. An area imaged in the early morning, for example, may look very different from one taken at the hottest time of the day, not to mention how different it looks from a VNIR image of the same area. This can make matching features against reference imagery difficult.
Two TIR reference images taken at different times.
TIR radiometric challenges
Temperature control
TIR sensors measure surface temperature, and therefore temperature control for these sensors is particularly important. The sensors are also extremely sensitive to temperature changes. This can result in blurry images and distortion of radiometric values.
A blurred TIR image due to temperature range outside of the ideal operating range.
Horizontal and vertical streaking
In TIR imagery, horizontal and vertical streaking is common. Thermal imagery inherently carries a lower signal, pushing values closer to the noise floor. This makes imagery more sensitive to noise and calibration more difficult.
Image with pronounced vertical streaking due to non-uniformity correction issues.
Lack of reference data
It is important to calibrate the sensor radiometrically to ensure that reported pixel measurements are correct. Measurements from your TIR sensors must be compared to reference data where the temperature is physically measured on the ground, to establish absolute ground truth. Without a reliable way to check TIR temperature values against ground truth, you cannot fully trust the temperatures reported by the satellite. This influences applications like wildfire detection or crop stress monitoring.
To recap
TIR data is difficult to calibrate and process because of a lack of suitable reference data sets. Coarser-resolution imagery, lower signal-to-noise ratios, and sensitivity to temperature also make it more difficult to process.
