Can Satellite-Observed Wildfire Incidents Alone Project Next-Year State Transition? A Case Study in British Columbia, Canada Using a Physics-Regularized Conditional Categorical Generative Model

We introduce the first AI model that predicts wildfire spread with the placement of both permanent and temporary firebreaks. Our spatiotemporal model learns from simulation data to capture how fire interacts with changing suppression efforts over time. Our model runs fast enough for near real-time use and performs well across different wildfire events. This approach could lead to better tools for helping decision-makers understand where and when firebreaks are most effective.