Driving wildfire spread prediction by multi-source real-time observations

Abstract

Wildfire spread prediction based on data assimilation (DA) enhances forecast accuracy through observational data integration, which essentially achieves optimal parameter estimation by minimizing discrepancies between observed and predicted fireline positions. However, DA performance shows strong sensitivity to fireline observation errors. To enhance DA effectiveness in the presence of inevitable observational errors, this study establishes a novel DA framework driven by multi-source observational data:(1) Based on the assumption that the observational errors in the x-y coordinate system follow a normal distribution, a method for estimating the confidence interval of the fire line vertex position was established; (2) A multi-source data fusion method is established, and an uncertainty quantification method for fused fireline positions is developed using probability theory; (3) The weighted root mean square error (RMSE) is implemented as the fitness function in parameter estimation, through which the Differential Evolution (DE) algorithm is guided by vertex-specific weights derived from uncertainty analysis for optimal parameter identification. The methodology is validated through both large-scale controlled experiments (spanning 10,000 m${}^2$ with coordinated UAV and watchtower monitoring) and simulation studies. Results demonstrate significant improved prediction accuracy compared to single-source DA approaches.