Firebird: Autonomous UAV Detection and Monitoring of Wildfires

Wildfires are becoming an increasing concern as climate change reduces rainfall, causes earlier snow melt, and increases lightning storms. According to a 2023 study by NASA's JPL, wildfires created more CO2 than all other man-made sources in Canada combined. This only increases the effects of climate change.

The majority of these fires are in northern Canada where the only organized detection method is satellites. Current satellite systems in Canada (FIRMS, NOAA), have 3 key limitations:

They can only reliably detect fires that are > 50 hectares (although in perfect conditions detection's of fires of around 5-10 hectares are possible.)

They have massive 6 - 12 hour gaps when no satellite is overhead

They are limited by cloud cover and atmospheric attenuation of thermal signatures.

As northern Canada is largely unpopulated these fires can smolder for days or even weeks before being detected, by then suppression is no longer an option and at best they can only be contained. The best way to limit these fires is early detection; a small smoldering fire < 1 m2 is much easier to put out than a raging fire that covers several km2. This project aims to offer a more reliable, cost-effective, and easy-to-deploy alternative to current wildfire detection and monitoring methods, while being orders of magnitude more sensitive.

This project took a different approach to current detection methods and designed a long range, autonomous drone equipped with a thermal camera to detect wildfires at a much higher resolution and accuracy than current methods.

There are two challenges that this project overcomes to create a viable wildfire detection method:

Designing and testing an Unmanned Aerial Vehicle (UAV) capable of flying hundreds of kilometers completely autonomously in order to cover the vast area of northern Canada. This required a radically different approach to conventional UAV airframes, with maximum efficiency being the primary design constraint. With the only other consideration being enough space and power to carry the modest payload of a Raspberry Pi and thermal camera.

Developing and testing a custom system to reliably detect and mark wildfires using their thermal signature. A Raspberry Pi 5 and a TOPDON TC001 thermal camera were used in combination with a custom detection algorithm to detect hot-spots and further evaluate whether or not they were a wildfire. This needed to be highly accurate with few false positives, while still not sacrificing sensitivity. The benchmark was to be able to detect smoldering fires less than 1 m2, even when partially obscured by the tree canopy.

Unlike current satellite systems, this integration of a long-range autonomous drone with high-sensitivity thermal detection eliminates coverage gaps, removes dependence on weather conditions, and can detect fires orders of magnitude smaller than existing methods.

Airframe:

A radical concept was used in the development of the UAVs airframe. It has been known for a while that it is possible to significantly increase the efficiency of an aircraft by moving the center of gravity over the neutral point, completely eliminating trim drag. But this causes the aircraft to be dynamically unstable and thus unflyable. However, with modern advancements in computer stabilization it is now possible for a computer to control the aircraft, increasing its efficiency by as much as 20%.

The airframe went through over 25 different iterations, with extensive CFD testing being done on each iteration using open source CFD models Flow5 and XFLR5. The result was an airframe with an impressive Lift/Drag ratio of 24.5, while having an optimal cruising speed of 70km/h. For reference, a high efficiency airliner operating at much higher Reynolds numbers has a Lift/Drag ratio of only ~ 16 - 19.

Detection:

An algorithm was designed to detect hot-spots using the TOPDON TC001 thermal camera. Two different algorithms were run in tandem and the results were compared in order to maximize detection reliability and sensitivity. The first was a basic thresholding algorithm, the second was a more sophisticated anomaly detection algorithm that took the overall temperature of the surrounding terrain and detected anomalies.

The detection system was field tested in real world conditions. A small controlled fire was lit partially obscured by a tree, and the UAV flew a pattern over the fire at 3 different altitudes: 120m, 100m, and 70m. This resulted in an average detection accuracy of 96% when using both algorithms, 74% when using each algorithm headlessly, and an accuracy of 100% across the board when operating at an altitude of less than 100m. Zero false positives were observed in over 500 detections.

This project has designed a long range drone, capable of autonomously detecting wildfires before they have a chance to reach a dangerous size. This has significantly improved on current early detection and wildfire monitoring methods. It is more than 1,000x the sensitivity of satellites such as MODIS. A fleet of ~1000 UAVs can cover all high risk forests in Canada daily at a fraction of the cost of current methods.

Limitations

While these are very promising results, there are still several limitations to be addressed:

Airframe:

No real-world data of the unstable flying wing has been collected as the prototype was irreparable damaged on its first test flight due to an autopilot malfunction. As yet it remains a proof of concept backed by significant CFD testing. All detection tests were conducted using a more conventional UAV frame, and all range and area coverage estimates are based on this limited airframe. It is predicted that the improved airframe will more than double both range and area covered in one day.

Detection:

While the current method is highly reliable, it is possible that an increase in false positive will occur during hot summer days when the ground is heated up by the sun. However, the current system is only a stepping stone for data gathering and is soon to be replaced by a ML model

Weather:

While still proven to handle winds of > 70km/h, weather is still a limiting factor; though no more so than satellites due to cloud-cover.

This project has designed and tested a viable wildfire early detection and monitoring system, but there are still several areas that need improving before this can be deployed to its full potential in the real-world.

Improved Airframe

The current airframe, while still quite capable, does not have range and flight duration to effectively cover the massive areas needed.

ML Model:

A Machine learning model is being developed that will further improve detection sensitivity and reduce false positives.

Multi Sensor Fusion

Integration with CO2, visible light sensors are being researched that will further increase detection accuracy and swath width.

There are many people who have given me support and advice in this project:

Dr. Fereydoon Diba and the people at Fleming College for their helpful ideas and for giving me access to some equipment I needed.

The team at Trent University for giving me this opportunity to participate in CWSF and for funding and organizing my trip.

My siblings for helping me with documenting my project by taking photos.

My parents who have given me support throughout and have been with me the whole way. Without you this would not have been possible.

[1] Natural Resources Canada. (n.d.). Canadian Wildland Fire Information System. Canadian Forest Service, Government of Canada. https://cwfis.cfs.nrcan.gc.ca/home

[2] Canadian Space Agency. (2026, April 29). About WildFireSat. Government of Canada. https://www.asc-csa.gc.ca/eng/satellites/wildfiresat/about.asp

[3] Environment and Climate Change Canada. (2026, March 10). Greenhouse gas emissions projections. Government of Canada. https://www.canada.ca/en/environment-climate-change/services/climate-change/greenhouse-gas-emissions/projections.html

[4] Huang, M. S., & Wichmann, B. (2026, March 12). Machine learning estimates on the impacts of detection times on wildfire suppression costs. PLOS ONE, 19(11), e0313200. https://doi.org/10.1371/journal.pone.0313200

[5] Johnston, J. M., Johnston, L. M., Wooster, M. J., Brookes, A., McFayden, C., & Cantin, A. S. (2026 April 10). Satellite detection limitations of sub-canopy smouldering wildfires in the North American boreal forest. Fire, 1(2), 28. https://doi.org/10.3390/fire1020028

[6] NASA Earth Observatory. (2026, October 25). Tracking Canada's extreme 2023 fire season. NASA. https://earthobservatory.nasa.gov/images/151985/tracking-canadas-extreme-2023-fire-season

[7] NASA Fire Information for Resource Management System (FIRMS). (2026, March 5). Ultra real-time detection. NASA Earthdata Wiki. https://wiki.earthdata.nasa.gov/spaces/FIRMS/blog/2022/07/14/258343755/

[8] NASA Jet Propulsion Laboratory. (2026, April 20). New NASA study tallies carbon emissions from massive Canadian fires. https://www.jpl.nasa.gov/news/new-nasa-study-tallies-carbon-emissions-from-massive-canadian-fires/

[9] Thompson, D. K., Fusina, G., & Jackson, P. (2026, April 10). Evaluation of ground-based smoke sensors for wildfire detection and monitoring in Canada. Fire, 9(4), 141. https://doi.org/10.3390/fire9040141

[10] Wente, M. (2026, March 10). Staffed lookout towers aren't relics from the past — they're key to wildfire detection. The Globe and Mail. https://www.theglobeandmail.com/opinion/article-staffed-lookout-towers-arent-relics-from-the-past-theyre-key-to/

IMAGES

[11] O'Kane, C. (2026, April 20). See pictures and videos of the 2023 Canadian wildfires and their impact across the planet. CBS News. https://www.cbsnews.com/news/pictures-videos-canadian-wildfires-footage-photos-2023/

[12] Maps.com. (2026, April 20). Canadian wildfires: Animated map shows spread of wildfire and smoke. https://www.maps.com/canadian-wildfires-animated-map-shows-spread-of-wildfire-and-smoke/