People sit on an Okanagan Lake beach viewing a distant wildfire.
How wildfires spread is more variable and unpredictable than Canada’s standard models assume, new research from UBC Okanagan data scientists shows.
Ladan Tazik, lead author of a new study in Fire and UBC Okanagan doctoral student, used advanced computer vision tools to capture fire behaviour with a level of detail that wasn’t possible even a few years ago.
Her work sheds light on the random elements of fire movement—information that could reshape how fire behaviour is modelled and forecasted in an era of worsening wildfire seasons.
“Image processing techniques let us quantify fire behaviour in real time, including the parts that don’t follow consistent patterns,” says Tazik. “By capturing the randomness in how fires spread, we can build models that better reflect reality and help improve decision-making during active fire events.”
Tazik led the design, analysis and modelling that form the backbone of the study.
She used the “Segment Anything Model”, a state-of-the-art AI tool, to extract fire perimeters from experimental burn videos frame by frame to study fire spread dynamics.
This allowed her to study directional fire spread on sloped terrain without assuming the fire behaves predictably or spreads in a simple line.
Her analysis confirmed something firefighters may know instinctively: fires race uphill. But when she compared her measurements with the values used in Canada’s official Fire Behaviour Prediction System, the numbers didn’t always line up.
Real fires often moved faster, and the influence of slope wasn’t consistent from place to place.
She tested the method on ponderosa pine and Douglas fir fuels often used in fire research.
This highlights that small differences in fuel, wind and terrain can add to the unpredictability of fire and introduce important variations in how it spreads.
Even under nearly identical conditions, the flames didn’t behave the same way twice.
In practical terms, that means most fire spread is shaped by randomness—far more than today’s deterministic models capture.
“These results show that we need to pair every spread estimate with a measure of uncertainty,” Tazik explains. “Simply multiplying by a slope factor isn’t enough. Fire is dynamic, and our models should acknowledge that.”
Research supervisor Dr. W. John Braun says the project demonstrates how emerging computer vision tools can transform wildfire science.
“Tazik proposed innovative ways to tackle this difficult modelling problem,” he says. “Her work shows how high-resolution perimeter data and advanced modelling can help us understand the real variability in fire behaviour. That’s essential if we want to move toward more probabilistic, data-driven prediction systems.”
The study also included contributions from Dr. John R.J. Thompson, Assistant Professor of Data Science, Mathematics and Statistics, as well as other partners who provided the experimental and field video datasets.
While the fuel experiments supported the research, Tazik alone led the segmentation and modelling components.
Tazik says the next step is to expand the approach to more fuel types and fire conditions and use airborne or satellite imagery to study fire spread dynamics.
With more Earth observation and remote sensing tools available, she sees an opportunity to build models that better capture wildfire dynamics while embracing the inherent uncertainty of fire, rather than smoothing it away.
“Fires don’t behave perfectly,” she says. “Our tools shouldn’t pretend they do.”
