A thermal drone does not measure “heat” the way a thermometer does. It measures infrared radiation, and then software, physics and human judgment decide whether that signal means a failing solar cell, a thirsty crop or a missing hiker. This review explains the whole measurement chain, not just the gadget.
- Why thermal drones matter
- How the technology evolved
- Platforms and sensors
- Radiometric vs non-radiometric
- Flight planning and acquisition
- Calibration and error sources
- Processing, mapping and AI
- Applications compared
- Comparing studies fairly
- Challenges and research gaps
- Future directions
- FAQs and conclusion
1. Introduction: Seeing What Ordinary Cameras Cannot
Temperature changes usually appear before damage becomes visible. A crop under water stress closes its stomata and its canopy warms. A loose electrical connection resists current and runs hotter than its neighbors. A person lying in a dark field radiates more heat than the cool ground around them. A normal camera records none of this, because it only captures reflected visible light.
Putting a thermal camera on a UAV adds three things at once: height, speed and repeatability. Surveys that once needed hours of walking, or a lift and scaffolding, can be flown in minutes. But a thermal drone is better understood as a measurement system than as a flying camera. Its output depends on the aircraft, the sensor, the weather, the target surface and the processing chain together.
The NETD and accuracy figures are typical manufacturer datasheet values for uncooled cores and vary by product, so replace them with your payload’s own datasheet. The irradiance figure is commonly quoted from IEC TS 62446-3; confirm the exact wording in the standard. A manufacturer accuracy figure describes test conditions, not what you will get in the field.
What this review sets out to do
- Explain how thermal cameras on UAVs actually sense and report temperature.
- Show how flight planning, calibration, weather and processing change data quality.
- Compare applications honestly, including where thermal results mislead.
- Assess the real contribution of AI, sensor fusion and autonomy.
- Point out research gaps that still hold the field back.
2. From Aerial Photography to Temperature-Based Observation
2.1 How thermal UAVs evolved
Two trends met in the middle. Drones became cheap, stable and easy to automate. At the same time, uncooled microbolometer detectors shrank in size, weight and power, which removed the need for the heavy cooled cameras once restricted to manned aircraft and defense work. Early drone work centered on visible photography and mapping. Thermal payloads came next, often as qualitative “heat spotting” tools. Today’s workflows include georeferenced thermal orthomosaics, time-series comparison, automated anomaly detection and, increasingly, fused multi-sensor payloads.
Stage 1: Heat spotting
Live video, human operator, visual hot spots. Fast and useful, but no traceable temperature values.
Stage 2: Thermal mapping
Radiometric images stitched into georeferenced maps for area-wide analysis and repeat surveys.
Stage 3: Intelligent monitoring
AI-assisted detection, fused sensors and flight paths that react to what is found.
2.2 What makes a thermal drone different from an RGB drone?
RGB and thermal answer different questions, so they complement rather than replace each other.
| Aspect | RGB camera | Thermal camera |
|---|---|---|
| Detects | Reflected visible light | Emitted and reflected infrared radiation |
| Best at | Cracks, dirt, shape, color, labels | Temperature contrast, hidden heat sources |
| Works in darkness | No | Yes (passive sensing) |
| Typical resolution | Many megapixels | Often 0.05–0.3 megapixel on drone payloads |
| Main weakness | Blind to heat | Cannot see through glass and is easily misled by reflections |
3. Technical Foundations: The Drone and the Thermal Camera as One System
3.1 Choosing the UAV platform
Multirotor
- Hovers and orbits structures
- Multiple viewing angles
- Short endurance, often 20–40 minutes
Best for: roofs, towers, PV strings, rescue.
Fixed-wing
- Efficient over large areas
- Longer endurance
- Cannot hover; needs launch and landing space
Best for: farms, forests, corridors.
Hybrid VTOL
- Vertical takeoff plus efficient cruise
- Heavier and costlier
- More complex to maintain
Best for: large sites with little launch space.
3.2 How a thermal camera senses temperature
The detector never reads surface temperature directly. It receives a mix of radiation emitted by the target, radiation reflected from surroundings (sky, sun, nearby buildings) and radiation altered by the atmosphere between target and lens. Software then converts that signal into a temperature using assumed values. If the assumptions are wrong, the number is wrong, even when the picture looks sharp. Manufacturer documentation lists the parameters that matter: emissivity, reflected temperature, relative humidity, atmospheric temperature and distance. Emissivity is the most important to set correctly, and when emissivity is low, reflected temperature becomes important too; humidity, air temperature and distance matter more at longer ranges.
| Band | Wavelength | Typical detector | Where it fits |
|---|---|---|---|
| LWIR | about 8–14 µm | Uncooled microbolometer (VOx or amorphous silicon) | Most commercial UAV payloads; ambient-temperature targets, agriculture, buildings, PV |
| MWIR | about 3–5 µm | Usually cooled (for example InSb) | Specialist work such as gas imaging and very hot targets; heavier and costly |
3.3 Specifications that actually matter
Resolution gets the headlines, but it is only one piece. Judge a payload on the combination below.
| Parameter | What it controls | Common mistake |
|---|---|---|
| Detector resolution (e.g. 336×256, 640×512) | Pixels on target | Assuming more pixels guarantees accurate temperatures |
| Pixel pitch and lens focal length | Ground sample distance (GSD) | Ignoring the lens and comparing detector size only |
| Thermal sensitivity (NETD) | Smallest detectable temperature difference | Confusing sensitivity with absolute accuracy |
| Accuracy | Closeness to a reference temperature | Trusting the spec outside the stated test conditions |
| Frame rate | Motion handling, video use | Using a slow-rate unit for moving targets such as animals |
| Field of view | Coverage per image | Choosing a wide lens for small-target detection |
| Radiometric output | Whether temperature data are saved | Buying a visualization-only unit for quantitative work |
| Weight and power | Flight time, gimbal compatibility | Overloading the airframe |
Worked example. Ground sample distance can be estimated as pixel pitch × altitude ÷ focal length.
4. Radiometric vs Non-Radiometric Thermal Imaging
This is the single most overlooked distinction when reading papers or buying equipment. A non-radiometric camera produces a colored picture designed for viewing. A radiometric camera keeps per-pixel data that can be converted to temperature after the flight.
Non-radiometric
- Good for live search and rough triage
- Contrast is often auto-stretched, so colors change from image to image
- Poor for comparing surveys over time
Radiometric
- Supports temperature extraction and correction
- Allows emissivity and reflected temperature to be set afterwards
- Required for trend analysis and defensible reports
5. From Flight to Insight: Data Acquisition
5.1 Flight planning decisions that change the data
- Altitude: lower flights give finer GSD and less atmospheric path, but cover less ground and take longer.
- Speed: microbolometers have a response time, so fast flight can blur small features.
- Viewing angle: oblique views raise reflection effects, especially on glass and glossy panels. Near-nadir is often safer for PV; modest tilt can help on façades.
- Overlap: thermal images contain little texture, so stitching often needs higher overlap than RGB mapping.
- Time of day: pre-dawn suits wildlife and building envelopes; solar noon suits PV under load; early afternoon is common for crop water stress.
- Weather: wind cools surfaces, cloud shifts irradiance during a flight, and rain or fog attenuates the signal.
5.2 Best acquisition window by application
| Application | Usual timing | Reason |
|---|---|---|
| Solar PV | Clear sky, high irradiance, modules under load | Faults only show when current and sunlight are present |
| Building envelope | Pre-dawn or evening with a strong indoor–outdoor difference | Reduces solar loading and exposes heat loss |
| Crop water stress | Midday to early afternoon, stable sun | Canopy temperature differences are largest |
| Wildlife | Pre-dawn or after dark | Animals contrast strongly with cool ground and vegetation |
| Search and rescue | Night or low-sun hours | Less heat stored in rocks, roads and soil |
5.3 Field checklist you can use before every survey
Before leaving
- Confirm the camera is radiometric and which temperature range or gain mode it uses.
- Check lens, field of view and expected GSD at your planned altitude.
- Choose a time window for your application (see table above).
- Check airspace rules, permissions and weather forecast (wind, cloud, rain).
- Print or load the target emissivity assumptions.
On site
- Power the camera on early and let it stabilize before the first reading.
- Record air temperature, humidity, wind and cloud cover.
- Measure reflected apparent temperature, for example with crumpled aluminum foil as a reflector.
- Estimate emissivity by comparing a surface of known tape with the camera reading.
- Place at least three ground reference targets or contact sensors across the expected temperature range.
In flight
- Fly at the altitude that gives enough pixels across your smallest target.
- Keep speed low enough to avoid smearing; use higher overlap than for RGB mapping.
- Prefer near-nadir views for glossy surfaces such as PV glass, and log the angle.
- Repeat a reference target at the start and end of the flight to check drift.
- Pause or stop if wind or cloud changes the scene noticeably.
After landing
- Compare camera readings with reference targets and note the bias.
- Apply emissivity, reflected temperature, humidity and distance corrections to radiometric files.
- Extract temperatures from individual frames; use the mosaic mainly for location.
- Verify every flagged anomaly with a second method before reporting a fault.
- Save the log: settings, weather, targets, software versions.
6. Calibration and Temperature Reliability
A thermal image can look clean and still be quantitatively wrong. Reliability depends on controlling each error source or at least recording it.
| Error source | Effect on the result | Practical control |
|---|---|---|
| Wrong emissivity | Systematic temperature offset | Use material tables, then verify with contact probe or reference tape |
| Reflected apparent temperature | False hot or cold areas, especially on shiny surfaces | Measure it on site; change viewing angle |
| Atmospheric transmission | Loss of signal with distance and humidity | Enter distance, humidity, air temperature; fly lower |
| Detector drift and warm-up | Slow offset during a flight | Allow warm-up, use shutter or non-uniformity correction, repeat reference checks |
| Wind and solar loading | Surfaces change temperature during the survey | Log weather; fly in stable conditions; avoid mixing flights from different times |
| Mixed pixels | Small targets averaged with background | Lower altitude, longer lens |
A reliable workflow records camera settings, weather, calibration steps, reference measurements and uncertainty assumptions. Where possible, check a sample of thermal readings against independent ground measurements such as contact sensors, calibrated targets or blackbody references.
7. Processing, Mapping and AI-Assisted Interpretation
7.1 The processing chain
Figure 1. Conceptual workflow of a drone-based thermal imaging system. This is a proposed structure, not an experimental result.
Raw thermal frames usually need correction, band alignment, noise reduction and georeferencing. Stitching them into an orthomosaic makes locations easy to share, but it also introduces risk: blending and interpolation can smooth a small hot spot or create seams that look like real thermal edges. For quantitative work, many analysts extract values from individual radiometric frames and use the mosaic mainly for location.
7.2 What AI does well, and where it breaks
Good uses
- Locating panels, roofs or animals in large image sets
- Flagging candidate anomalies for human review
- Segmenting plots, canopy or modules
- Detecting change between repeat flights
Common failure points
- Small or imbalanced training sets
- Labels made from the image itself instead of ground truth
- Poor transfer to new sensors, seasons or sites
- High accuracy scores that hide false alarms from warm rocks, roofs or reflections
8. Application Domains: Where Thermal Drones Make a Difference
8.1 Precision agriculture
Plants cool themselves by transpiration. When water is limited, stomata close and canopy temperature rises relative to well-watered plants and to air temperature. Indices such as the Crop Water Stress Index use that difference to map irrigation needs. The difficulty is that wind, vapor pressure deficit, canopy structure and soil background all influence temperature, so thermal data work best when paired with weather data, multispectral imagery and reference plots. A review of UAV thermal work in precision agriculture lists plant water stress, disease detection, yield estimation and phenotyping as the main uses, and stresses that correct interpretation rests on understanding thermal radiation.
8.2 Infrastructure and building diagnostics
Drones can survey roofs, façades, pipelines and industrial plants quickly. Typical clues include missing insulation, thermal bridges, moisture and abnormal equipment heating. A thermal pattern is a lead, not a verdict: moisture, air leakage and material change can look alike, so follow-up with contact tests or moisture meters is standard good practice.
8.3 Solar PV and electrical inspection
Thermal UAV surveys can scan large solar plants far faster than a walking inspector. Typical findings include hot cells, bypass diode faults, string-level mismatches and connector heating. Industry guidance such as IEC TS 62446-3 covers outdoor thermography of PV modules and stresses adequate irradiance and consistent viewing conditions. Recent experiments back this up: a Danish study tested irradiance from 200 to 1000 W/m² and drone heights of 8 to 20 m, and found altitude mattered little compared with irradiance in that range; a Korean study found the angle between camera and panel affects measured temperature. Ground resolution matters here: a single hot cell needs enough pixels across it, or it will be averaged away.
8.4 Disaster management and search and rescue
Thermal imaging helps responders scan wildfire perimeters, collapsed structures and wilderness at night. Detection drops with dense canopy, glass, thick smoke particulates in some conditions, warm backgrounds and small target size. Coordination, communication links and privacy also shape real deployments.
8.5 Environmental and wildlife monitoring
Thermal drones have been used to count large mammals and nesting birds, locate animals in dense terrain and study habitat use. Detection depends on body size, canopy, ambient temperature and altitude. Flights must avoid disturbing animals, and raw counts should not be treated as true population size without correcting for missed animals. In a test on captive white-tailed deer of known number, a thermal drone estimated about 78% of the true density, rising to 92% for evening flights when thermal contrast was better.
8.6 Applications at a glance
| Domain | What thermal reveals | Main benefit | Main limitation |
|---|---|---|---|
| Agriculture | Canopy temperature, water stress patterns | Targeted irrigation and field variability maps | Weather and canopy structure confound readings |
| Buildings | Heat loss, moisture suspects | Fast roof and façade screening | Needs a temperature difference; reflections and sun load |
| Solar PV | Hot cells, string faults | Rapid large-site screening | Needs irradiance and sufficient resolution |
| Rescue and fire | People, hot zones | Night and low-visibility search | False detections, canopy cover |
| Wildlife | Animals against cool background | Less intrusive than ground surveys | Missed detections; disturbance risk |
9. Comparing Existing Research Fairly
A review adds value by explaining why studies disagree. Differences in sensors, altitude, weather, target type, labels and metrics often matter more than differences in algorithms. A fair comparison table should record the items below for every study.
| Column | What to extract |
|---|---|
| Platform and sensor | UAV type, thermal core, resolution, lens, radiometric or not |
| Target and setting | Application, site, season, background |
| Acquisition | Altitude, speed, angle, time, weather |
| Calibration and processing | Emissivity handling, corrections, mosaicking method |
| Analysis method | Manual, threshold-based, machine learning, deep learning |
| Validation | Reference sensors, ground truth, independent test sites |
| Metrics | RMSE or bias for temperature; precision, recall for detection |
| Limitations | Stated weaknesses and reproducibility of the workflow |
Selected published studies
The table below summarizes six peer-reviewed or published items found in journals and conference proceedings. It is based on their published abstracts and summaries. Read the full text before you cite any number, and note that the studies differ in target, sensor and metric, so their results should not be ranked against each other.
| Study | Application | Platform and sensor (as reported) | Calibration and validation | Key finding | Limitation to note |
|---|---|---|---|---|---|
| Kelly et al. 2019 | Temperature accuracy of UAV thermal cameras | Low-cost, uncalibrated uncooled TIR camera, tested in lab and in flight | Empirical-line calibration against ground calibration points | At least three ground points were enough to convert output to temperature; sensor-temperature dependency, wind and drift degraded results | Results tied to the cameras tested; may not transfer to other models |
| Messina and Modica 2020 | Precision agriculture | Review across many studies | Not a single experiment | Covers water stress, disease, yield estimation and phenotyping; says correct use needs knowledge of thermal radiation | A secondary source; check original studies |
| Aghaei et al. 2018 | PV plant fault inspection after storm damage | Lightweight infrared camera on a DJI Phantom 3 Advanced | Field inspection of a 3 MWp plant in South Brazil | Detected hot spots, defective bypass diodes and short- or open-circuited strings, with damage noticeable from 40 m | Conference paper at one site; not a controlled accuracy test |
| Shin et al. 2024 | PV hot spots and camera incidence angle | UAV TIR camera (check paper for model) | Multiple regression of temperature against humidity, emissivity, height, wind, irradiance, ambient temperature and incidence angle | Incidence angle was analyzed as an extra variable; the regression explained 63.6% of variability | More than a third of variability is unexplained; single plant with 20° tilt |
| DTU study 2025 | PV defects under varying irradiance and altitude | Drone infrared thermography (check paper for model) | Known defective and soiled modules in two operating strings; 200–1000 W/m²; 8, 10, 14 and 20 m | Altitude effect generally negligible compared with irradiance; soiling cases read 5–10 °C lower at higher altitude because of lower resolution | Altitude tested only up to 20 m, at one plant |
| Beaver et al. 2020 | Wildlife population estimation | Miniaturized thermal sensor on a drone | Captive deer population of known size (151–163 animals) | Density estimate was 78% of the known value, and 92% for evening flights | Captive enclosure; free-ranging, forested sites may give lower detection |
Section-level takeaway: the strongest evidence is about conditions (irradiance, time of day, sensor temperature, viewing angle), not about which camera or algorithm is best.
Results should be grouped by type of evidence: qualitative anomaly detection, temperature estimation, classification performance and validated field outcomes. Numbers from different groups, datasets or metrics should not be ranked against each other.
10. Persistent Challenges and Research Gaps
Measurement reliability
Environmental effects and inconsistent calibration reduce comparability between surveys.
Resolution limits
Small or distant targets occupy too few pixels for reliable detection or temperature estimates.
Endurance
Battery capacity, payload weight and wind restrict coverage per flight.
Interoperability
Proprietary file formats and inconsistent protocols block data sharing and re-analysis.
AI generalization
Models trained on one sensor or region often drop in accuracy elsewhere.
Validation
Many studies show convincing images without independent validation or uncertainty analysis.
Regulation and ethics
Airspace rules, privacy and data security influence where and how flights can happen.
11. Future Research Directions
| Direction | What it would change |
|---|---|
| Multisensor fusion (thermal + RGB, multispectral, LiDAR) | Adds geometry and context that explain thermal patterns |
| Uncertainty-aware thermal maps | Reports confidence next to each temperature and detection |
| Edge AI onboard the drone | Faster decisions with less data sent over weak links |
| Autonomous re-inspection | Drone revisits a suspect area at a better angle or lower altitude |
| Standard benchmarks and protocols | Makes algorithms and sensors comparable across studies |
| Cross-sensor generalization | Models that survive new cameras, seasons and sites |
| Digital twins and predictive maintenance | Links thermal history to asset records and failure risk |
| Energy-efficient operations | Longer missions and smarter route planning |
12. Frequently Asked Questions
Can a thermal drone see through walls or glass?
No. Ordinary window glass blocks LWIR, so the camera sees the glass surface and reflections on it. It can reveal heat conducted to a wall’s outer surface, but not what is behind it.
Is a higher-resolution thermal camera always more accurate?
No. Resolution helps you detect and outline small targets. Accuracy depends on calibration, emissivity settings, atmosphere and the camera’s own specification.
Do I need a radiometric camera?
For quick searching or rough triage, a non-radiometric unit can be enough. For temperature trends, reports or comparison between flights, choose radiometric.
Can thermal drones work in rain or fog?
Performance drops. Water droplets absorb and scatter infrared and wet surfaces change emissivity and cooling. Many operators avoid thermal inspection in rain.
Can AI replace the human thermographer?
Not yet for high-stakes decisions. AI is useful for screening large datasets, while trained people are still needed to rule out reflections, confirm causes and sign off findings.
Do I need permission to fly a thermal drone?
Rules depend on country and use case. Thermal flights are subject to the same airspace, licensing and privacy rules as other drone flights, and some uses such as flights over people or beyond visual line of sight need extra approval. Check your local aviation authority.
Conclusion
Drone-based thermal imaging is a powerful way to find temperature patterns across farms, buildings, solar plants, wildlife habitats and disaster zones. But the value of a thermal UAV system lies in the entire measurement chain: sensor choice, flight planning, calibration, processing, validation and interpretation.
The most useful next steps for the field are traceable measurements, reproducible methods, robust and well-tested AI, shared benchmarks and tighter links to real decisions. Until then, treat a thermal image as a strong clue, then confirm it.
References
- Kelly, J., Kljun, N., Olsson, P.-O., Mihai, L., et al. (2019). Challenges and best practices for deriving temperature data from an uncalibrated UAV thermal infrared camera. Remote Sensing, 11(5), 567. https://www.mdpi.com/2072-4292/11/5/567
- Teledyne FLIR. Measuring temperatures: measurement parameters (emissivity, reflected temperature, relative humidity, atmospheric temperature, distance). docs.flir.com
- FLIR Systems. Thermographic measurement techniques (emissivity, reflected apparent temperature, distance, relative humidity). support.flir.com
- IEC TS 62446-3: Photovoltaic (PV) systems, requirements for testing, documentation and maintenance, Part 3: Photovoltaic modules and plants, outdoor infrared thermography. IEC Webstore
- Technical University of Denmark (2025). Influence of irradiance and drone altitude in infrared thermography inspections of photovoltaic plants. EPJ Photovoltaics, 16, 31. https://doi.org/10.1051/epjpv/2025020
- Shin, H., Khoshelham, K., Lee, K., Jung, S., Kim, D., Lee, W. (2024). Effect of incidence angle on temperature measurement of solar panel with unmanned aerial vehicle-based thermal infrared camera. Remote Sensing, 16(9), 1607. https://doi.org/10.3390/rs16091607
- Aghaei, M., Madukanya, E. U., de Oliveira, A. K. V., Rüther, R. (2018). Fault inspection by aerial infrared thermography in a PV plant after a meteorological tsunami. CBENS 2018. https://doi.org/10.59627/cbens.2018.203
- Messina, G., Modica, G. (2020). Applications of UAV thermal imagery in precision agriculture: state of the art and future research outlook. Remote Sensing, 12(9), 1491. https://doi.org/10.3390/rs12091491
- Beaver, J. T., Baldwin, R. W., et al. (2020). Evaluating the use of drones equipped with thermal sensors as an effective method for estimating wildlife. Wildlife Society Bulletin, 44(2), 434–443. https://doi.org/10.1002/wsb.1090
Reference list checked against search results on the publication date of this draft. Check each source’s current page, and the IEC standard’s exact wording, before final publication.
Before you publish as a journal review
- Add a documented systematic literature search: databases, search strings, date range and inclusion criteria.
- Build the comparison table from original papers, not from earlier reviews.
- Write a short research-gap synthesis that names what remains unresolved and which standards or experiments would settle it.
No Comment! Be the first one.