A Neural Network Model of Wildfire Spread for Arbitrary Topography, Fuel Load Distribution, and Wind Field

Due to climate change and other factors, wildfires have become more destructive and widespread. In January 2025, multiple fires ignited in Los Angeles, destroying 15,946 structures, damaging 1,953 structures, and killing 28 people. Until we can prevent wildfire ignitions, one step we can take to mitigate the damage of wildfires is to predict how wildfires will spread. This foresight will allow emergency responders to arrange firefighting resources and set backing fires. In the past several years, researchers have used neural networks to predict wildfire spread in order to achieve low latency inferences. However, current published works have several limitations, such as finite input domains, poor lookahead times, and the inability to incorporate external weather predictions. In this research, we have created a GAN that can overcome these limitations. Given any land topography, initial fuel distribution, and wind field, our GAN can predict the burn scar at some future time. Our method also accounts for arbitrary ignition locations. Because our model is trained on simulated data, we can generate as much data as we wish, and we can push the lookahead time as far out as we wish. This means our methodology can be used to learn input features in infinite dimensions and predict wildfire spread far enough into the future to be useful. After training on only 1588 training points, our GAN performed well on the validation set, predicting the correct burn scar with an intersection over union of 0.90. The inference time was only 0.140 second.