U of A researcher using AI to get a step ahead of wildfires
New tools could map out complex weather data to help firefighters predict where forest fires are likely to break out.
By BEV BETKOWSKI
Weather and fuel—two leading wildfire culprits—are now in the crosshairs of a University of Alberta researcher hoping to use machine learning against them.
By leveraging artificial intelligence (AI) to sort and map reams of meteorological data, Mike Flannigan and co-researchers from the Canadian Forest Service and University of Waterloo want to better predict where forest fires could break out and take hold so firefighting agencies can plan ahead.
“We want to complement existing decision-making tools so they can make the best calls possible in dealing with fire to help protect communities,” Flannigan said.
- RELATED: New AI program fights fire with data
- RELATED: U of A gets funding to build cube satellite to help fight wildfires
By pinpointing where wildfires are likely to take root, fire crews and helicopters can be efficiently deployed, Flannigan said. That advance warning is crucial when it can take from three to seven days to get outside resources to the fire, he noted.
Fire managers would be able to predict severe fire weather in a particular area, determine whether the needed resources are available and then plan to have them on hand.
“For instance, if it’s wet in Quebec, crews and planes can be sent to Alberta to deal with wildfires there,” he said.
Flannigan is exploring the potential of neural network software that processes weather patterns and variables like temperature, pressure, humidity and wind speed to create far more detailed maps than currently exist.
“Think of a baby who sees a human face: they start to distinguish ears, nose and eyes, and as they distinguish further, they can see even more detail like whiskers and sideburns,” Flannigan said. Neural networks work similarly, to identify severe weather patterns in multiple layers of existing weather data.
“It could be turned into a map that would identify vulnerable areas and what times the fire weather would be severe,” he added.
The researchers are also interested in working with fire managers to develop an application for remotely sensing and mapping data about fuel layers in the forest.
“When there’s a fire on the landscape, we want to know what fuel it’s burning into, like grass, conifers and aspen. Fire management agencies do have fuel maps, but they’re often spatially coarse and we want to go to a much finer resolution with a lot more detail.”
Through remote sensing, maps can also be updated more often. Currently, it could be 18 to 20 years before an area that’s been mapped is assessed again, with many changes happening in between, Flannigan noted.
Updated, richly detailed spatial maps can give fire managers a better feel for how fuel for the fire is structured vertically and horizontally.
“In a high-intensity wildfire, a map with this application can help identify what we call ladder fuels like understory shrubs (plants that grow under tree canopies), which allow the fire to spread into the treetops,” Flannigan explained.
“The current one-size-fits-all system in use doesn’t tell us that; it can only make assumptions about the understory. But this new kind of machine learning will have much more detailed information about the fuel structure.”
The applications will go through pilot studies over the next two years to see whether machine learning can indeed improve on traditional methods. Flannigan, who started exploring its potential with one of his undergraduate students in 2016, believes it can.
“It’s not a panacea, but for certain problems with large volumes of data to process, machine learning allows us to see relationships that aren’t always obvious using traditional methods and approaches. We think machine learning approaches for wildfire management show a lot of promise.”