A wildfire forecasting system powered by artificial intelligence was around 30% better at identifying dangerous fire conditions in advance than the models currently used, according to new research.

The Australian and New Zealand researchers’ study, published in the International Journal of Wildland Fire, trained a machine learning model on historical weather and fire data from three different fire-climate regimes across Australia, specifically the subtropical Sunshine Coast, peri-urban Brisbane and temperate Hobart areas. They then tested it against Australia’s current wildfire risk system, the fuel-specific Fire Behaviour Index.

The machine-learning model outperformed the current forecasting model by 10% to 30%, depending on the scenario, the researchers found. The AI-based model also doubled potential financial savings compared with the Fire Behaviour Index.

An immediate limitation was the frequency with which the current model updated wildfire data compared to the new machine learning model. Current systems could only compute wildfire forecasts hourly and release a risk rating to the public only once a day. But the machine-learning system could update its information every 30 minutes, providing near-real-time insight.

“Wildfires don’t wait for the daily forecast. If fire weather conditions can change within hours, our warning systems need to move just as quickly,” lead author Alberto Ardid said in a Te Whare Wānanga o Waitaha | University of Canterbury press release. We originally developed the model as a proof of concept in one region. However, in this new study, we tested it across multiple regions to see whether the approach works under different fire weather conditions.”

Click here to read the full research.