High-fidelity CNN-based surrogate modeling for wildfire spread forecasting.

The rising frequency of large-scale, destructive wildfires has significantly affected not only natural ecosystems but also critical infrastructure, human lives, and properties. Given the devastating and often irreversible environmental and financial consequences, researchers from diverse fields are actively seeking solutions to improve resilience against wildfires. Fire suppression, one of the most effective strategies to mitigate wildfire damage, relies heavily on rapid and high-fidelity forecasting of fire spread. These predictions are essential for planning evacuations by state or local emergency management agencies, implementing preemptive de-energization strategies for electric utilities, and coordinating fire containment efforts by firefighting teams. However, a significant bottleneck across all these planning processes is the significant computational burden imposed by high-resolution wildfire modeling, the demand for improved predictive accuracy, and the need to integrate diverse and large-scale datasets. Since response time is crucial for wildfire risk management, this paper proposes a deep learning-based surrogate model to predict fire spread in just a fraction of a second. We developed and trained a convolutional neural network (CNN) model that efficiently predicts wildfire propagation. The proposed model demonstrates high efficiency, achieving an F1 score of 0.92. The contributions of this paper are twofold: (1) a fast, high-resolution CNN model that can support wildfire-related public safety power shutoff (PSPS) planning for electric utilities, and (2) a practical tool for firefighting and evacuation teams to support rapid and data-informed risk assessment.