Optimal Transport for Variational Data Assimilation
Abstract. Variational data assimilation methods are designed to estimate an unknown initial condition of a model using observations. To do so, one needs to compare model outputs and observations. This is generally performed using Euclidean distances. This paper investigates another distance choice: the Wasserstein distance, stemming from optimal transport theory. We develop a variational data assimilation method using this distance. We investigate the impact of the scalar product, the 5 gradient choice as well as the minimization algorithm. With appropriate choices, we show successful results on preliminary experiments. Optimal-transport-based optimization seems to be promising to preserve the geometrical properties of the estimated initial condition.