Updated: Specs from manufacturer pagesNo paid placements

How thermal drones work

Thermal drones work by focusing the infrared heat given off by everything in view onto a grid of tiny sensors called a microbolometer. Each sensor warms slightly, changes its electrical resistance, and that change becomes one pixel of the image sent to your controller.

How do thermal drones work?

Scene
Heat leaves every surface as long-wave infrared
Lens
Germanium or chalcogenide glass focuses it
Microbolometer
Each pixel warms and changes resistance
Processor
Corrects, measures, applies a palette
Video link
Sends the image to the controller
Screen
Pilot sees heat, picks a spot to measure
Diagram: how a thermal drone turns heat into a picture on your controller.

The drone part is ordinary: GPS, motors, a gimbal that holds the camera steady, and a radio link. What makes it thermal is the camera. Its lens can’t be glass, because glass blocks long-wave infrared, so thermal lenses use germanium or chalcogenide glass. Behind the lens sits the sensor.

The thermal imaging camera sensor: an uncooled microbolometer

Every current drone thermal camera uses an uncooled microbolometer: a chip covered in tiny suspended membranes, one per pixel, made of vanadium oxide (VOx) or amorphous silicon. Infrared warms each membrane by a fraction of a degree, its resistance changes, and the electronics read that change. “Uncooled” means it works at normal temperature. Military cameras that cool their sensors to cryogenic temperatures are far more sensitive and far more expensive, and none are sold on consumer or small enterprise drones.

The pixel size (pixel pitch) is usually 12 µm on current sensors. DJI quotes 12 µm for both the Matrice 4T and the Mavic 3T. Smaller pitch means a smaller, lighter camera for the same resolution.

DJI uncooled microbolometer: what DJI publishes

DJI describes its drone thermal cameras as “uncooled VOx microbolometer” sensors at 640×512 with 12 µm pitch, and publishes NETD as “≤50 mK @ F1.0” for the Matrice 4T. The “@ F1.0” matters: NETD depends on the lens aperture, and a figure without one can’t be compared fairly. DJI makes these modules itself (its lineup is on DJI thermal drones); Teledyne FLIR supplies the Boson cores in Skydio and Parrot drones.

Infrared camera sensor specs, decoded

  • Resolution (for example 640×512) is the pixel count. Four times the pixels of 320×256 means a target can be identified from roughly twice the distance.
  • NETD (noise-equivalent temperature difference), in millikelvin (mK), is the smallest temperature step the sensor can separate from its own noise. Lower is better: 30 mK beats 50 mK. It matters most in low-contrast scenes like dawn or summer nights.
  • Frame rate. Most drone thermal cameras run at 30 Hz. Some versions are capped at 9 Hz or below, because US export rules treat thermal cameras faster than 9 Hz as more sensitive technology. The Parrot ANAFI USA runs at 9 fps.
  • Radiometric means the camera records a temperature for every pixel (saved as R-JPEG), not just a picture. Inspectors need it.
  • MSX is Teledyne FLIR’s name for blending edges from the visual camera into the thermal image so it’s easier to read.

Thermal sensor resolution, drawn to scale

Same scene width, same warm figure. Each square is one thermal pixel (grid lines hidden at 1280 so it stays legible). More pixels on a target means you can tell a deer from a hog, or a person from a rock, from higher up.

320×25681,920 pixels. Parrot ANAFI USA. A person at 100 m spans roughly a quarter as many pixels as on a 640 sensor.
640×512327,680 pixels. Every other drone here. The current standard for recovery, rescue and inspection.
1280×10241,310,720 pixels. DJI’s Super Resolution output on the Matrice 4T is processed from a 640×512 sensor, not native.

Grids are drawn at one-tenth scale: 32 columns stand for 320, and so on.

Thermal sensor specs for every current pick; each value links to the manufacturer page it came from.
DroneResolutionNETD (lower is more sensitive)Frame rate
Matrice 4T640×512 (1280×1024 in Super Resolution mode)source≤50 mK @ f/1.0source30 Hzsource
Matrice 4TD640×512source≤50 mK @ f/1.0source30 Hzsource
EVO Lite 640T640×512sourceAutel does not publish NETD for this modelNot published by the manufacturer
EVO II Dual 640T V3640×512sourceAutel does not publish NETD on its US pages; reviewers cite ≤40 mK, which we have not been able to confirm30 Hzsource
EVO Max 4T V2640×512sourceAutel does not publish NETD for this modelNot published by the manufacturer
Skydio X10640×512source<30 mKsourceNot published by the manufacturer
ANAFI USA320×256 (photo); 1280×720 upscaled videosource≤60 mK (stated as thermal sensitivity)source9 fpssource

NETD (noise-equivalent temperature difference) is the smallest temperature step the sensor can separate from its own noise, in millikelvin (mK). 30 mK beats 50 mK. How thermal sensors work.

Emissivity: why the camera can be fooled

A thermal camera measures the infrared a surface gives off, then calculates a temperature using that surface’s emissivity: how well it radiates heat, on a scale from 0 to 1. Skin, water, brick and most paints are around 0.9 or higher and read accurately. Polished metal can be below 0.1, so it mostly reflects its surroundings. A shiny roof vent can look freezing on a hot day because it’s mirroring the cold sky. Radiometric drones let you set emissivity, and some let you change it afterwards in software.

Gain modes and temperature range

Drone thermal cameras usually have two gain modes. High gain covers a narrow range (DJI quotes −20 to 150 °C on the Matrice 4T) with finer detail, for people, animals and buildings. Low gain covers a wide range (up to 550 °C on the Matrice 4T) for fires and hot industrial equipment, with less sensitivity.

The click: flat-field correction

Every few minutes an uncooled thermal camera briefly freezes and you may hear a click. That’s a small shutter closing so the camera can recalibrate each pixel against a uniform surface (flat-field correction, or non-uniformity correction). Without it, the image would slowly drift and develop fixed patterns. It’s normal, and you can usually trigger it manually before an important capture.

Thermal crossover

Crossover is the main reason drone deer recovery happens after dark. Twice a day, around sunrise and sunset, the ground and the objects on it can pass through the same temperature. During that “crossover” the thermal image goes flat and targets nearly vanish. Night and early morning give the best contrast for finding people and animals.

How many thermal pixels land on a deer or a person?

Resolution alone doesn’t tell you what you’ll see. What matters is how many thermal pixels cover the target, which depends on the lens and how high you fly. We calculated it for the drones whose makers publish the thermal camera’s field of view, assuming the camera points straight down over flat ground.

Pixels along the target’s length × width. Each pixel’s footprint on the ground is shown in brackets.
Drone (thermal DFOV)Deer at 200 ftPerson lying down at 200 ftDeer at 400 ftPerson lying down at 400 ft
Skydio X10 (41° source)29 × 9 px [5.2 cm]32 × 9 px [5.2 cm]14 × 4 px [10.5 cm]16 × 4 px [10.5 cm]
DJI Matrice 4T (45° source)26 × 8 px [5.7 cm]30 × 8 px [5.7 cm]13 × 4 px [11.5 cm]15 × 4 px [11.5 cm]
DJI Mavic 3 Thermal (61° source)19 × 6 px [7.8 cm]22 × 6 px [7.8 cm]10 × 3 px [15.6 cm]11 × 3 px [15.6 cm]

How to read it: a common rule of thumb from thermal-camera “DRI” criteria is that about 1–2 pixels across a target tells you something warm is there, around 6 tells you what kind of thing it is, and around 12 lets you identify it with confidence. A deer that’s only 3–4 pixels wide at 400 ft is easy to detect but hard to tell from a coyote, which is why pilots spot from height, then drop lower or switch to the zoom camera to confirm. Autel and Parrot don’t publish thermal field of view, so we can’t calculate theirs.

Software for thermal images

Radiometric images (R-JPEG) need software to read temperatures afterwards. DJI Thermal Analysis Tool is free for DJI files; Teledyne FLIR offers its own tools for Boson-based cameras; mapping packages can stitch thermal images into temperature maps for roofs and solar sites.

Why thermal images look soft

A 640×512 thermal image has about a third of a megapixel; the visual camera beside it has 48. That’s physics and cost, not a fault. Long-wave infrared has wavelengths around 20 times longer than visible light, so lenses and pixels are bigger. Processing such as DJI’s Super Resolution (1280×1024 output from a 640 sensor) sharpens the result but doesn’t add sensor pixels.

Already own a drone? See whether you can fit a thermal camera for a drone you have. Ready to choose? The thermal drone buying guide turns these specs into a decision, and the best thermal drones list applies them.

Questions people ask

Why does my thermal camera click and freeze?

That’s flat-field correction: a small shutter closes for a moment so the camera can recalibrate every pixel. It’s normal and keeps the image from drifting.

What does mK mean on a thermal camera?

Millikelvin, a thousandth of a degree. A sensor rated ≤50 mK can separate temperature differences of about 0.05 °C from its own noise; ≤30 mK is more sensitive.

Do thermal cameras need to warm up?

Briefly. Images appear at once, but radiometric readings stabilise after a few minutes, so let the camera run before you measure.

Can a thermal camera see thermal updrafts?

Not directly. Air is nearly transparent to long-wave infrared, so a thermal camera sees the sun-warmed field or car park that feeds an updraft, not the rising air itself.