Identifying critical fire spread to the wildland–urban interface using cellular automata and reinforcement learning

Wildfires threatening the wildland–urban interface present significant risks to community safety, especially under conditions of inadequate vegetation management and adverse weather. Accurately identifying scenarios in which fire reaches this interface is crucial for timely evacuation planning and risk mitigation. This study presents a computational method using cellular automata and stochastic simulations to model wildfire spread. Stochastic scenarios generated through the cellular automata are employed to train a reinforcement learning model, which leverages computer vision techniques to interpret multiple layers representing diverse environmental factors. This enables the reinforcement learning agent to identify and prioritise critical fire trajectories that could impact the wildland–urban interface. The framework adapts the Rothermel surface fire spread model within a cellular automata structure, providing a simplified yet effective simulation of fire propagation under variable conditions. The proposed approach was validated using synthetic and real-world case studies, demonstrating its potential for integration with geographic information systems. Results suggest this approach enhances the identification of critical fire spread scenarios and improves computational efficiency for real-time applications. By enabling real-time recognition of high-risk events, our framework supports more informed evacuation strategies and fire management decisions around the wildland–urban interface.