A Probabilistic Approach to Wildfire Spread Prediction Using a Denoising Diffusion Surrogate Model

Abstract

Thanks to recent advances in generative AI, computers can now simulate realistic and complex natural processes. We apply this capability to predict how wildfires spread, a task made difficult by the unpredictable nature of fire and the variety of environmental conditions it depends on. In this study, We present the first denoising diffusion model for predicting wildfire spread, a new kind of AI framework that learns to simulate fires not just as one fixed outcome, but as a range of possible scenarios. By doing so, it accounts for the inherent uncertainty of wildfire dynamics, a feature that traditional models typically fail to represent. Unlike deterministic approaches that generate a single prediction, our model produces ensembles of forecasts that reflect physically meaningful distributions of where fire might go next. This technology could help us develop smarter, faster, and more reliable tools for anticipating wildfire behavior, aiding decision-makers in fire risk assessment and response planning.

Figures (10)

• Figure 1: Deterministic and stochastic models.
• Figure 2: (a) Data collection, including canopy density, canopy cover, landscape slope and local wind speed corresponding to the forest area affected by the Chimney fire, California, in 2016; (b) Possible directions considered for each cell when simulating fire propagation using cellular automata (CA). (c-e) CA simulated wildfire spread samples from random ignition points at intervals of $20$ hours.
• Figure 3: (a) A snapshot (frame) of a sample of wildfire spread simulated with CA. (b) Grayscaled snapshot. (c) A complete wildfire spread simulation sequence is recorded as a sample where each snapshot is a frame. (d) An example illustrating the creation of the ensemble testing dataset. Each sample in the dataset is generated with a different initial ignition position and consists of a sequence of input-target frame pairs. The CA model is executed multiple times starting from a single frame $\mathbf{x}_n$, producing multiple simulated versions of the next frame $\mathbf{x}_{n+1}$. These simulated frames are then averaged to produce an ensemble simulation $\bar{\mathbf{x}}_{n+1}$, representing the probability of fire spread at the next time step. Each pair of frames $(\mathbf{x}_n,\bar{\mathbf{x}}_{n+1})$ forms an input-target pair, providing data for evaluating the model's ability to predict the probabilistic transition between consecutive wildfire states.
• Figure 4: Diffusion model training process
• Figure 5: (a) Attention Res-UNet architecture; (b) residual block architecture; (c) attention block architecture