As frontline firefighters battle the worst wildfire season in British Columbia’s history, the pilot project has been rolling out at a furious pace at the British Columbia Wildfire Service headquarters in Kamloops.
Director of Forecasting Services Neil McLoughlin and his team have spent much of 2023 implementing two artificial intelligence programs that analyze the Kamloops and Coastal Fire Centres and take on the work typically done by human analysts: data inputs on geography, weather forecasts, drought conditions and even detailed descriptions of the types of vegetation where new fires are emerging.
“Typically, it takes two to four hours for each fire simulation (when done by humans), and we can run as many as 14 in a single day,” he told CTV News in an exclusive interview. “Last year, we were simulating at least 200 fires a day, with some of them being repeated as many as twice in one day as new weather forecasts came in.”
That means decision-makers won’t have to wait to assess what equipment and personnel to deploy where, which could make a big difference when responding to explosive wildfire outbreaks like those that hit B.C. last year.
The entire state was subject to the AI analysis in March, at a time when cooler, wetter weather was slowing the fire season in most areas. This allowed the team to pay close attention to a small number of fires and evaluate how well the parallel program was predicting what would happen as new fires started.
The Canadian software, FireCast, and American Wildfire Analyst, usually provide similar results, but sometimes give conflicting predictions about where and how quickly a fire will spread. For now, the team plans to run both systems and compare the results.
Machine learning can save you valuable time
Prolonged drought and thunderstorms have sparked hundreds of wildfires in recent years, some of which have destroyed infrastructure and homes.
Computer modeling has been around for years, but when there are hundreds of fires burning at the same time and limited resources to deploy, the manpower needed to predict the spread of a fire simply can’t keep up. Having a digital helper that can perform simple tasks quickly allows for faster, more informed deployment.
“This takes a lot of the work out of the equation and allows us to get ahead of the fires,” McLoughlin said. “The impetus for this effort is to turn back the clock and give us more time to think about what a fire is going to bring and how to use our resources effectively, and respond to fires before they happen.”
For the time being, BCWS will keep the results of its machine learning internally, as it does not rely solely on modeling but considers multiple factors before deciding how to fight a wildfire (or whether to fight it in remote areas). As with all modeling, the emphasis is on predictions representing what is likely to happen, not what will certainly happen.
Case Study
The AI program analyzed numbers and conditions surrounding the Horsethief Creek fire in the Kootenay region, which began burning in late July last year, prompting evacuation orders and restricting access to risk areas.
Their predictions had the fire coming dangerously close to Invermere, which contributed to the BCWS’s decision to fight the fire from the east to avoid the worst of it. After weeks of heavy attack, the fire was eventually contained to a small fraction of the area that modelling warned it could burn without firefighting intervention.
The two systems work on very limited data sets, which differs from large language models like ChatGPT and provides more reliable results for very specific goals. This also means that BCWS can customize what you want to be part of the analysis, which we will do gradually over the coming years.
McLoughlin stressed that humans still make the decisions about what to do. “This doesn’t replace humans, but it will provide more timely information to decision makers and allow for greater coverage of the state during major fire seasons like we have in 2023.”
