Artificial intelligence has for three years now served as a kind of gut check for Alberta Wildfire staff making decisions about how to allocate resources. 

Years ago, the provincial agency in charge of wildfire response in Alberta’s protected forests identified a potential use of AI: predicting where wildfires would break out the next day. 

“Their observation was that they spend money on trying to be preventative and a lot of the time, those resources just idle,” recalled Graham Erickson, a machine learning developer at AltaML, an Edmonton-based software company that has contracts in both the public and private sectors. 

Protected forests cover more than half of Alberta and are divided into 10 regions. Each region is overseen by an officer or manager whose responsibilities include suppression planning. 

“We call fire a science and an art – and this is kind of (the) art piece,” Olivia Aftergood told CTV News Edmonton during an interview in late September. 

“There’s a lot of things that go through your mind when you’re in those types of scenarios.”

Today, AltaML’s creation looks like a digital dashboard: a map of Alberta’s forest regions is colour coded according to what it has calculated as the likelihood of a wildfire breaking out the next day. 

The day that Aftergood spoke to CTV News Edmonton, all forests were coloured blue – denoting low risk of a fire starting the next morning – except the Peace River region in northwestern Alberta, where the AI had concluded there was a higher afternoon fire risk. That area was coloured yellow. 

An image of the wildfire-predicting artificial intelligence program built by AltaML for Alberta Wildfire. (Jeremy Thompson / CTV News Edmonton)

The program uses the same information available to Alberta Wildfire staff, but it differs from the province’s assessment of fire danger, defined as a measure of “the ease of ignition, rate of spread, difficulty of control and fire impact.”

“If a fire happened, how bad would it be?” Erickson said. “Which is a slightly different question from, ‘What is the chance of a fire occurring?'”

The AI program has been fed and trained with about a decade of fire, weather and ecological data. It also ingests daily weather modelling provided by provincial meteorologists and considers factors like whether it’s a long weekend – when human recreation is more likely to cause a fire. It presents the level of risk as a probability of a fire occurring. 

According to Erickson, the program has a 50-per cent success rate of “catching” fires, meaning it predicts about half of all fires that happen. Its track record is better when looking at the accuracy of its predictions: it is correct 80 per cent of the time it says a fire will happen. 

Aftergood stressed that the AI does not “take over” decision making, but supports it. 

“As we have people retire who have that extensive knowledge … and we’re getting new people in – they don’t necessarily have that experience. And sometimes that lack of experience can be really stressful for some people,” she said.

“This tool is really meant to help them on their journey as they’re getting more comfortable in those fire management roles, just because it is such a big responsibility and you’re covering such a big land base.”

Erickson added, “That’s the idea with, actually, a lot of our approaches to (machine learning): Human intuition is very powerful. What we’re looking at with machine learning is a way of quantifying some of those feeling-based judgments.” 

He believes there are two benefits to this. 

“One, it’s really hard to make disciplined, reproducible operations around gut feeling. But if you have some numerics that back up those gut feelings, now you can put them into process and have paperwork and audits … The other thing about that is gut feeling isn’t something you have Day One on a job. So a 20-year plus duty officer is going to have really great gut feeling, but a brand-new duty officer might not have that same sort of intuition.”

Originally built in three months and piloted in 2022, the product has gone through numerous renovations since then with feedback of the wildfire experts and continues to do so. 

According to Aftergood, most duty officers and fire managers use the AI while making suppression plans. 

Every year, fed with the data of the previous wildfire season, the tool becomes better, she said, even when the previous year was an anomaly like 2023 was. To date, wildfires this year have burned 708,540 hectares of forest, according to the Alberta Wildfire Status Dashboard, potentially the third-highest amount since 2019. In 2023, more than 2.2 million hectares of forest burned. 

Erickson believes this application of AI is among the technology’s best uses. With the City of Calgary, AltaML developed a program that predicts when rainwater catch basins will need maintenance to get ahead of resident requests. 

But, he said, an effective program requires a “rich” dataset, like the decades of data collected by Alberta Wildfire. 

“I’m not interested in technology in a bubble,” he told CTV News Edmonton. “I don’t think that technology is here just because it’s interesting; I think it’s because it’s useful.”

The legislated wildfire season, which started early this year in February, isn’t over until the end of October. On Thursday, 54 fires were burning in Alberta’s protected forests, all of which were classified as under control or being held. 

With files from CTV News Edmonton’s Galen McDougall and Jeremy Thompson



Source link

author-sign