The camera is a commodity. The calibration is not.
Radiometric fixed cameras start around €1,400. What decides whether the data is usable is everything around the camera.
The stack
Four layers.
| Layer | What it is | Why it matters |
|---|---|---|
| Sensor | Radiometric thermal with per-pixel °C output, plus co-located air temperature and humidity. Fixed, handheld or UAV. | A non-radiometric camera gives a colour picture, not data. No °C, no delta-T, no trend, no product. |
| Reference | An on-frame reference surface where the use case needs absolute accuracy. Emissivity, distance and angle recorded per installation. | Field-grade cameras are typically ±2 °C. Several published disease deltas are smaller than that. Without a reference you measure your own noise. |
| Edge | Inference on site. Segmentation, tracking, delta computation, threshold and trend logic. | Latency, bandwidth and privacy. Video does not leave the farm. |
| Decisions | Alerts routed by role — maintenance, agronomy, herd. Ranked by estimated cost, not by anomaly count. | An alert nobody owns is noise with a timestamp. |
Calibration protocol
What we do that a camera on a pole does not.
-
Ambient compensation
Air temperature and humidity measured at the camera, not at the farm gate. A ±1 °C error becomes a 28–83% CWSI error.
-
Emissivity per surface
Leaf 0.94–0.96. Reference plate about 0.97. Galvanised steel is so low and inconsistent that silo walls had to be painted before thermal imaging worked at all in a published trial.
-
Fixed geometry
Distance and angle recorded and enforced. In cattle hoof studies, view angle alone shifts readings by about 1 °C.
-
Measurement windows
Canopy indices near solar noon under clear sky. Animal imaging after acclimatisation, out of sun and wind.
Wind at or above 7 km/h lowered cattle eye temperature by 0.43 ± 0.13 °C; at or above 12 km/h, by 0.78 ± 0.33 °C.
-
Commissioning baseline
Every asset gets a reference image at install. Trends run against it, not against a manufacturer’s spec sheet.
Sources: Irrigation Science 2022 (CWSI error); ASABE stored-grain silo trial (emissivity); Schaefer et al. 2007, Res Vet Sci 83:376 (wind); Animals 2025, 15(8):1086 review (view angle).
Hardware position
We are camera-agnostic, and we will tell you what to buy.
We do not manufacture cameras. We integrate them. Typical list bands, observed August 2026:
| Class | Indicative price |
|---|---|
| Fixed radiometric, Ethernet or GigE | €1,400 – €10,500 |
| Handheld inspection | €500 – €15,000 |
| UAV thermal payload | €6,000 – €11,000 |
Prices drift. The specification that does not drift is radiometric output. Ask any vendor for a per-pixel temperature file, not a screenshot.
Three grades. Published next to every claim.
Thermal imaging in agriculture ranges from independently replicated science to claims with no supporting study. Most vendors present both the same way. We do not.
Two different things carry this grade, and the source line always says which. Either the finding is independently replicated in peer-reviewed literature — forehead temperature tracking THI at r = 0.808 in dairy cattle, N = 488, supported by a 2026 review of 83 articles.
Or it is mature practice in an adjacent industry, such as electrical inspection thermography, where the physics is settled and the evidence is operational.
We will never present the second kind as if it were the first.
Real published results from small samples, single sites, or lab conditions not yet reproduced in a working barn or field. The sample size is always attached.
Example: over 81% accuracy detecting digital dermatitis on the day of first clinical signs — 38 animals, one farm, one study.
Mechanistically plausible, actively researched, not yet supported by field evidence we would stake a claim on.
We run these as research, at research pricing. We do not put a performance number on the website.
What we refuse
Four things we will not sell you.
Oestrus detection by thermal camera
A 2023 review of ten primary studies found most report qualitative conclusions rather than standardised sensitivity and specificity.
One secondary source we could not trace puts specificity as low as 29%. We flag it as unverified rather than drop it, because even the possibility of that many false alarms should change your decision.
Activity collars and boluses do this job better and cheaper.
Mastitis detection as a product
The literature is split. Reported sensitivity ranges from 95.6% in a 62-cow study to 71–79% in a 155-cow field study.
Two further studies conclude infrared has low diagnostic ability, or fails at early detection. We run this as research. We do not invoice for it as a feature.
Disease diagnosis
Thermal screening for foot-and-mouth or African swine fever is a triage aid at most.
The only ASF-specific study we found had ten animals and calculated no sensitivity or specificity. The World Organisation for Animal Health’s own ASF diagnostic guidance does not mention thermal screening.
Fever screening at the farm gate
We found no evidence of any kind that this works, has been piloted, or is deployed for livestock premises.
It appears to be an extrapolation from pandemic-era human screening.
Privacy position
Our models look at animals, plants and equipment. Not at people.
Video is processed on site and is not uploaded. Under the EU AI Act, a system scoped to livestock, crop and equipment condition monitoring is not an Annex III use case. Animals are not natural persons.
Systems that evaluate worker performance are a different matter, and we do not build those. Under KVKK, the Turkish authority has stated that generic justifications such as checking whether employees are working are not legitimate purposes for camera monitoring.
We scope our models accordingly, and it is written into the contract. This is our position and our design constraint. It is not legal advice.
Questions
Questions we are actually asked.
Do I need to replace my existing cameras?
No. Thermal is an additional sensor, and most deployments use a small number of well-placed units rather than one per existing position.
Does the video leave my site?
No. Inference runs on an edge unit in your building. Aggregated measurements and alerts sync out. Frames do not.
What internet connection do I need?
Enough for alerts and daily aggregates. A 4G router is sufficient for a single building.
What happens if the connection drops?
Inference continues. Alerts queue locally and flush when the link returns. The edge unit is the system of record, not the cloud.
Which cameras do you support?
Any camera with true radiometric per-pixel output over Ethernet or GigE, plus supported handheld and UAV payloads. We specify exact models in the assessment.
Do you monitor people?
No. Our models are scoped to animals, plants and equipment, and that scope is contractual.