Forest Fires in Canada: An Alternative Model Produces Risk Charts 806 times faster than the Reference Simulation

Where BurnP3+ requires days or weeks per region, the network estimates in seconds the likelihood of burning, intensity and rate of spread.

As fires increase, Canada needs frequently updated national fire risk maps to plan for fuel management, land use planning or evacuation routes. However, reference simulators like BurnP3+ rely on thousands of Monte-Carlo prints and can mobilize days or even weeks for a single regional unit.

Multitasking network driven nationally

IA researchers and fire specialists propose a deep learning model that mimics BurnP3+ outputs from fuel, relief, ignition probability and regional weather. At the same time, it estimates the likelihood of burning, the average intensity and the average rate of propagation, and retains the hazard classes used downstream. Tested on counterfactual scenarios proposed by experts, he reacts physically plausiblely.

In a representative region, the calculation is 806 times faster on processor and uses only 13% of simulator memory. The model also transfers without re-training to another regional study, with a slight refining sufficient to significantly improve its calibration.