Data Assimilation for Wildfire Spread Modeling with Conditional Generative Models

Wildfire behavior modeling involves highly complex systems that capture intricate interactions between the atmosphere, terrain, and fuel properties. While these advanced models can offer reliable short-term predictions, their accuracy tends to decline over extended periods, highlighting the importance of incorporating data assimilation techniques. Traditional data assimilation methods in wildfire modeling often rely on simplifying assumptions or provide a single-state estimate. However, Bayesian data assimilation, which incrementally refines the probability distribution of model states based on ongoing observations, offers a promising yet underexplored alternative. In this work, we develop an approach for performing Bayesian data assimilation utilizing conditional generative models. The approach utilizes the forward model along with an approximate observation operator to produce samples from the predicted state distribution and the corresponding measurements, thus providing data for training. Once trained, a conditional generative algorithm is used to transport samples from the predicted distribution to the filtered distribution, conditioned on a true measurement. We demonstrate the proposed approach using Lorenz 63, a three-­­­­variable nonlinear chaotic system widely used for testing data assimilation algorithms, and the coupled atmosphere-wildfire model WRF-SFIRE, with measurements from VIIRS active fire satellite data.