John Walter Acevedo Valencia, Institut für Mathematik der Universität Potsdam

Hybrid Approach to High Dimensional Non-Gaussian Data Assimilation

Sebastian Reich and Maria Reinhardt


Particle filters are able to correctly assimilate observational information under strongly non-Gaussian conditions and present asymptotic consistence in the large ensemble size limit. However, due to the curse of dimensionality, they typically collapse when utilized in high dimensional problems. In order to tackle this problem, we propose a hybrid methodology where the Ensemble Transform Kalman Filter (ETKF) and the Ensemble Transform Particle Filter (ETPF) are used sequentially. Furthermore, localization techniques are consistently applied to both filters, which is the key to the applicability to high dimensional systems. We demonstrate numerically that our approach operates robustly and outperforms both Kalman and particle filters at moderate ensemble sizes.