As wildfires spread faster and hit more communities across Canada, researchers at the University of Toronto are developing new technology that could give first responders a critical advantage: the ability to predict how a fire will evolve in the hours ahead.
Steven Waslander, a professor at the U of T Institute for Aerospace Studies (UTIAS) in the Faculty of Applied Science & Engineering, is working with fire research scientists at Natural Resources Canada and Simon Fraser University on TankerVision - a project aimed at creating more accurate, AI-powered fire prediction models.
Such tools could help firefighters make key decisions about directing resources and ordering evacuations.
The researchers have teamed up with government agencies and industry to use existing field equipment and technology to collect real-time data about the fires. Partners include BC Wildfire Service, Alberta Wildfire, Canadian Forest Service, aircraft operator Conair Aerial Firefighting and Voxelis, an AI-powered aerial wildfire mapping company.
Wildfire growth is currently predicted based on a handful of controlled burns measured in the 1980s and data on weather and forest conditions - all filtered through decades of human experience, says Waslander. But the ground is shifting, and the science needs to keep pace, he says.
"This method hasn't been adapted to current fire severity increases due to climate change and involved a great deal of extrapolation from a limited dataset," says Waslander.
"Our big idea here is to use AI and computer vision to capture unprecedented numbers of fires from the air to build more sophisticated, data-driven wildfire prediction models."
To do this, piloted firefighting aircraft have been fitted with high-resolution colour and infrared cameras and computers.
"Once the plane is airborne and reaches a certain speed, it begins scanning for smoke or flames and records only when a fire is visible," says Waslander. "All the storage is on board the aircraft so our partners can review the data before we start working on it to make sure we aren't logging any confidential information."
First, the AI performs what's called segmentation: examining the image and sorting each pixel into fire, smoke or ground. But smoke can cloud this process. With no clear edges, it drifts, thins and blends into the terrain, making it tricky to trace where it begins and ends.
It's a problem Waslander is familiar with as principal investigator of the WinTOR all-weather driving program , which develops autonomous systems that can navigate through snow or heavy rain.
"The methods we use are all related to autonomous driving techniques - it's the same labelling, object detection and segmentation tasks, but this time, the challenge is to find the fire boundary," says Waslander.
"Here, we use segmentation models trained on all kinds of other challenges for segmentation, and then we fine-tune and adapt them to the small amount of data we have available to us. We're collecting a lot of data from the fire perspective, but from the AI perspective, it's still a tiny data set."
Once the images are segmented, the system maps them onto 3D terrain by matching each recording with satellite maps. This step relies on expertise from field robotics, where aligning imagery with older or imperfect maps is a common challenge. The merged data reveals how the fire is moving over time.
The team will then combine these measurements with weather data and wind readings to build AI models capable of forecasting fire growth.
"Last season, we logged 50 fires, which is currently the highest number recorded in one season," says Waslander. "This year, we plan to expand the data collection efforts to three aircraft and hope to capture more than 100 wildfires - sufficient to start developing and validating our prediction models."
The team is now preparing hardware for this fire season and hopes to expand partnerships across the country, including in Ontario, to help transform how Canada responds to wildfires. Its first research paper, focused on segmentation, has been submitted, with a full-system paper and public data set planned for release in the fall.
Waslander says the project resonates deeply with students and researchers who have witnessed the increase in severe fires first-hand.
"I think everyone involved in the project really feels passionately about the direction we're going and see it as an opportunity to make a real impact on keeping us safe in the years to come as things get hotter," says Waslander. "The contributions to Canadian safety and preservation of our country are really motivating and it's nice to be able to think our research can actually contribute to this."
