Title:A Sampling Filter for Non-Gaussian Data Assimilation

Abstract:Data Assimilation in operational models like atmospheric or Ocean models is almost impossible without posing many assumptions due to the complication of the model that is usually very high-dimensional and also due to non-linearity of the observation operator used to map the state space to the measurement space. The current operational ensemble-based filters like Ensemble Kalman Filter (EnKF), and Maximum Likelihood Ensemble Filter (MLEF) usually fail in case of nonlinear observations and/or non-Gaussian distributions. In this paper, we propose a new ensemble-based data assimilation method, named "sampling filter". The filter is used to sample directly from the posterior distribution. The sampling strategy in the analysis phase of the filter follows a Hybrid Monte Carlo (HMC) approach that is, in principle, applicable in cases where Gaussian assumption is violated and also when observations are nonlinear. In this work, the posterior is assumed to be Gaussian as a first step but we used several observation operators with different levels of non-linearity and differentiability. Several experiments testing the potential and performance of the proposed algorithm and comparing its performance with both EnKF and MLEF using six different observation operators are conducted. The Lorenz-96 model is used and the results are compared to EnKF and MLEF. The results show that this approach is promising, mostly with the new developments proposed recently in HMC sampling literature.