Data-driven surrogate model with latent data assimilation: Application to wildfire forecasting

The large and catastrophic wildfires have been increasing across the globe in recent decade, highlighting importance of simulating forecasting fire dynamics near real-time. This is extremely challenging due to complexities physical models geographical features. Running physics-based simulations for wildfire events real-time are computationally expensive, if not infeasible. In this work, we develop test a novel data-model integration scheme progression forecasting, that combines Reduced-order modelling, recurrent neural networks (Long-Short-Term Memory), data assimilation, error covariance tuning. modelling machine learning surrogate model ensure efficiency proposed approach while assimilation enables system adjust simulation with observations. We applied algorithm simulate forecast three California from 2017 2020. deep-learning-based runs around 1000 times faster than Cellular Automata which used generate training data-sets. daily perimeters derived satellite observation as Latent Assimilation An tuning also performed reduced space estimate prior errors. evolution averaged relative root mean square (R-RMSE) shows reduce RMSE by about 50% considerably improves accuracy. As first attempt at order spread our exploratory work showed potential data-driven speed up various applications.