Localized recursive spatial-temporal state quantification method for data assimilation of wildfire spread simulation

Abstract

Data assimilation is a procedure to improve the state inference by assimilating the real-time observation data into dynamic systems, such as wildfire spread simulation. Various techniques are used for data assimilation, such as sequential Monte Carlo methods, also called particle filters. In the standard sequential Monte Carlo methods, the used number of particles is the same during the entire process and their convergence is not measured or characterized for runtime state estimation. Therefore, to guarantee the convergence, an abundance of particles is required so that they can be widely distributed in the state space to converge to the true posterior of the system state. In the application of wildfire spread simulation, the spatial states are large and the system behaves in a heterogeneous manner in different fire areas. The heterogeneous feature and the spatially and temporally dynamic behavior of the wildfire spread simulation cause the state inference uncertainty to be dynamically changed. In this paper, we propose the localized recursive spatial-temporal state quantification method to measure the convergence of particles at runtime and apply various approaches to improve the state inference. To show its effectiveness, we apply two different algorithms – the adaptively perturbing the localized state space algorithm and the adaptive particle filtering algorithm – to the localized recursive spatial-temporal state quantification method to improve the state estimation and enhance the performance respectively. The designed experiments are used to show the effectiveness of the adaptively perturbing the localized state space algorithm and the adaptive particle filtering algorithm in the improvement of the state estimation and the performance of data assimilation in wildfire spread simulation.