Phoenix: Wildfire Prevention using an Autonomous Bionic Morphing Ornithopter and AI Models

CWSF 2026 Video Script

Canada’s 2025 wildfire season was the second-worst in the nation’s history. In Manitoba alone, 432 wildfires consumed more than 2.1 million hectares of land.

These wildfires resulted in an increase in resource waste, smoke pollution, and the number of crops burned.

Traditional wildfire prevention methods are often inefficient or expensive, which motivated me to develop an innovative alternative. I designed a bionic morphing ornithopter integrated with AI to address this challenge.

The ornithopter features advanced capabilities including autonomous navigation, object detection, obstacle avoidance, infrared sensing, and temperature and humidity monitoring to identify and assess areas susceptible to wildfires. To support prevention efforts, I incorporated a natural fire-retardant solution using Aloe Vera, deployed through a custom payload delivery mechanism.

In addition, I developed an AI model using Prophet to predict fire intensity and a YOLO ML model to identify ignition sources along with their severity. To validate the effectiveness of the system, I conducted multiple tests, including ANOVA analysis, computational fluid dynamics simulations, and performance evaluations.

This project demonstrates a novel, cost-effective, and autonomous approach to wildfire prevention with strong potential for real-world application. Thank you for listening.

Background:

According to the Canadian Wildland Fire Information System, 6,127 fires burned 8,922,148 hectares in 2025. In Manitoba, wildfires scorched more than 2.1 million hectares, which was the worst in 30 years. This forced 32,000 evacuations amid two states of emergency, costing Manitoba more than 225 million dollars in loss, simply from wildfires.

Traditional wildfire prevention devices like fire alarms, suppression and sprinkler devices, and aerial drones struggle with vast remote areas, high costs, efficiency, and extensive human monitoring. This includes the inability to autonomously predict and prevent fire spread or identify ignition sources in real time.

Objective

To create a fully autonomous bionic ornithopter with morphing wings and tail that uses artificial intelligence (AI) to predict wildfire intensity, with features such as obstacle avoidance, motion sensing, temperature & humidity sensing, object detection, and a payload mechanism.

To assess the efficiency of the ornithopter using computational fluid dynamics (CFD), statistical analysis (ANOVA), and create a natural fire retardant out of aloe vera.

Hypotheses

If the ignition source size decreases or its colour lacks contrast with its background, the ML detection accuracy will decrease, regardless of its intensity.

As the thermal emission value of an object increases, detection accuracy will also increase.

If the AI model’s prediction interval decreases or training data includes more years of data, its prediction accuracy will increase.

If the altitude or weight of the natural fire retardants within the ornithopter increases, then the payload mechanism accuracy will decrease.

Ornithopter Construction:

Build the ornithopter by wiring electronic hardware such as the Matek F405 Wing V2 Board, ESC, 1250 Kv Motor, MicroAir Compass, then verify firmware for ArduPilot configuration. Then build an ornithopter fuselage through foamboard, and attach servo motors to the tail and wings to allow morphing, and construct the wings and tail through flame-resistant PVC and carbon fiber rods. Design and integrate a 3D-printed flapping gearbox for wing movement.

Autonomous Navigation & Obstacle Avoidance:

Integrate the Matek F405 Wing V2 Flight Controller with the MicroAir GPS to enable the ornithopter to receive location and telemetry data for self-navigation and stability. Then connect the Ultrasonic Sensor (HC-SR04) to Arduino Nano, Arduino Nano Shield, and Arduino Code for distance monitoring between the ornithopter and obstacles.

Payload Mechanism & Fire Retardant:

Construct a custom Payload Mechanism using a servo motor to control a hinge and allow Aloe Vera (Natural Fire Retardant) to be dropped in desired places.

Prophet AI model & YOLO Detection Model:

Download data from NASA's FIRMS from (2019–2024) for Canada. Configure the dataset using Phrophet AI structure, which then allows for fires across Canada to be predicted at any time, place, and intensity. Then the model's accuracy is calculated through MAE. Collect image data for ignition sources, then use a YOLOv8 architecture to design a real-time detection algorithm for ignition sources.

PIR Infrared Motion Sensor/ Temperature & Humidity Sensor:

Use a Passive Infrared Sensor (HC-SR501) to identify heat motion by changes in infrared light emitted, by integrating with Arduino Nano, Arduino Nano Shield, and Arduino Code. Construct a Temperature & Humidity Sensor with a DHT11 Sensor, Arduino UNO, and Arduino Code allowing for constant air quality data.

Computational Fluid Dynamics Testing:

Simulate Computational Fluid Dynamics testing through SlimWorks by generating the ornithopter's custom design to determine aerodynamic efficiency through lift/drag coefficients.

YOLO Detection Model:

The machine learning model showed strong precision at high confidence, reaching perfect precision of 1.0 at 0.702 confidence, meaning high-confidence predictions were always correct. However, the F1 score peaked at only 0.51 at 0.146 confidence, indicating a weak balance between precision and recall at lower confidence levels. Recall reached 1.0 at zero confidence, meaning the model captured nearly everything.

Training and validation losses decreased steadily across 100 epochs without divergence, suggesting the model learned effectively and did not overfit. The label distribution plot showed that bounding boxes across the nine (9) object classes were generally small to medium in size. The position distribution also showed that bounding boxes were often clustered, suggesting that broader placement diversity could improve detection performance.

The confusion matrix showed that Cooking Oil, Wooden Sticks, and Propane Tank were the strongest classes, with many correct predictions. In contrast, some classes were more frequently confused with similar categories, showing room for improvement through class-specific training. The normalized confusion matrix confirmed this pattern, with Cooking Oil, Wooden Sticks, and Propane Tank performing well, and Cigarette, Cigarette Lighter, and Match Box were often misclassified or treated as background.

Prophet AI Model:

For AI wildfire forecasting, analysis of 2019–2025 fire radiative power data showed that the Prophet model tracked the rolling monthly median very well but was less accurate for day-to-day predictions. Performance was stronger for earlier periods, such as 2019, but accuracy dropped in later years due to outliers and increasing uncertainties. Summer peaks were especially pronounced, and the model captured these seasonal trends well up to 2024 before uncertainty widened in 2025. Seasonal analysis also confirmed that wildfire activity is dominated during summer.

The AI prediction model improved substantially over time for both Canada and Manitoba. By 2027, absolute error dropped to about 5–7 units, while relative error decreased from around 70% in 2024 to roughly 20%. This suggests that later model versions, better data quality, or improved training produced stronger forecasting results, especially for Manitoba, where local accuracy was nearly as good as national accuracy.

Autonomous Ornithopter:

The ornithopter demonstrated that the additional temperature, humidity, and infrared sensors were accurate, and autonomous navigation tests showed that it could follow pre-programmed GPS waypoints accurately, even after detours, while maintaining self-stabilization. Computational Fluid Dynamics (CFD) testing also showed improved lift and stable drag behavior as the angle of attack changed, indicating high aerodynamic efficiency. The ornithopter is lightweight and low-cost as it weighs less than 700g and costs less than $225.

Statistical Analysis:

A Two-Way ANOVA indicated that drop height significantly affected first point distance, with a significant main effect and interaction between height and weight, but the weight was not significant (Bonferroni Correction was done). A Two-Way Anova indicated that the final landing distance and drop height remained significant, but the main effect of weight and the interaction between height and weight were not. These results show that height was the key factor influencing landing performance.

YOLO ML Model:

The YOLO ML model demonstrated strong pattern recognition (perfect precision at high confidence) but struggled with consistent object detection (F1=0.51, mAP=0.494). The ML model had high recall but low precision. Higher confidence reduces false positives but misses objects, while lower confidence catches more objects at the cost of accuracy.

Prophet AI Model:

The Prophet AI Model predicted with high accuracy where and when fires might happen across Canada and their FRP Values to determine the fire intensities, and how intense fires might occur in Canada and Manitoba. The model showed that seasonality also played a pivotal role in when wildfires might occur for early prevention.

Ornithopter:

The ornithopter proved that learning from nature, biomimicry, can lead the future and aerial devices can also be cheap, effective, and competent in comparison to industry and research drones. The lift and drag coefficients from CFD show that lift increases and drag remains nearly constant as the angle of attack changes. The ornithopter's additional features proved to be valuable features in wildfire prevention with accurate temperature, humidity, and infrared values.

Statistical Testing:

Overall, the experiment shows that drop height is the primary factor affecting both first-point and final landing accuracy, while payload weight alone does not reliably improve accuracy; effective optimization requires selecting suitable height/weight combinations, particularly keeping drop height low when precise landings are needed.

Future Work:

On-board YOLO preprocessing to enable human-caused ignition sources during flight rather than post-flight.

Integrating solar panels to battery grid to increase overall flight time and decrease in environmental impact.

Testing ornithopter in natural air bodies to determine the impact of outliers in the real-world such as temperature, obstacles, and distance.

Wing flapping mechanism testing will optimize angle-of-attack control and improve ornithopter aerodynamics for stable autonomous operation.

Mass-manufacturing the ornithopter to further reduce costs and improve wildfire prevention methods worldwide with this novel approach.

Advanced materials to enhance ornithopter's durability, speed, efficiency, and biomimicry of natural bird flight characteristics.

Acknowledgement:

I sincerely thank my parents for providing materials and invaluable guidance. I would also like to thank my friend, Ayomipo Oduntan for their help with printing the gearbox for my ornithopter when my 3D printer jammed halfway through. I would also like to thank my regional science fair MSSS and MSSS judges for organizing their amazing science symposium year over year, and CWSF and CWSF judges for valuing their time with me while I present my project. Finally, I would like to thank Mr. Jared Thorlaksson for providing me important electronic hardware pieces for my project.

References

Websites:

NASA. (2024, March 19). Tech today: NASA helps find where the wildfires are. NASA. https://www.nasa.gov/general/tech-today-nasa-helps-find-where-the-wildfires-are/

The Weather Channel. (2015, July 17). Wildfire seasons nearly 20 percent longer worldwide, thanks to climate change: Study. https://weather.com/science/environment/news/wildfireseasons-nearly-20-percent-longer

Wildfire information. Province of Manitoba. (n.d.). https://www.gov.mb.ca/wildfire/index.html

Earth Science Data Systems, N. (2024a, August 9). Wildfires. https://www.earthdata.nasa.gov/topics/human-dimensions/wildfires

Wildfires - Canadian Red Cross. (n.d.-b). https://www.redcross.ca/how-we-help/emergencies-and-disasters-in-canada/types-of-emergencies/wildfires

Science projects. (n.d.). https://www.sciencebuddies.org/science-fair-projects/science-projects

Books/Articles:

Canadian wildland fire information system / système canadien d’information sur les feux de végétation. (n.d.). https://cwfis.cfs.nrcan.gc.ca/en/fire-history

Ye, X., Ye, Y., Huang, X., & Onega, T. (2026, February 18). Wildfires and public health: A comprehensive review of human-centric studies. GeoHealth. https://pmc.ncbi.nlm.nih.gov/articles/PMC12914487/

Rubinstein, D. (2026, March 27). Researchers work to protect Canadians from wildfires. Challenge What’s Possible. https://challenge.carleton.ca/research-protect-canadians-wildfires/

MacCarthy, J., Richter, J., Tyukavina, S., & Harris, N. (2025, July 21). The latest data confirms: Forest fires are getting worse. World Resources Institute. https://www.wri.org/insights/global-trends-forest-fires

News Articles:

Wildfires in Canada – news, tracking and Air Quality. CTV News. (n.d.). https://www.ctvnews.ca/canada/wildfires/

CBC/Radio Canada. (2026, January 25). New Analysis Highlights Canada’s wildfire paradox: Fewer fires, Greater Destruction | CBC news. CBCnews. https://www.cbc.ca/news/canada/

canada-wildfires-fewer-fires-more-damage-study-9.7051171 Global News. (n.d.). Wildfires: News, videos & articles. Global News. https://globalnews.ca/tag/wildfires/

Softwares:

SlimWorks Ideal Simulation CFD

OnShape

SPSS

Python 3

Arduino

ArduPilot

MissionPlanner