A Comparison of Linear and Non-linear Data Assimilation Methods Using the NEMO Ocean Model

The assimilation behavior of the Equivalent Weights Particle Filter (EWPF) is compared to the widely used LETKF in a data assimilation application using an idealized configuration of the NEMO ocean model. The experiments show how the different filter methods behave when they are applied to a realistic ocean test case. The LETKF is an ensemble-based Kalman filter, which assumes Gaussian error distributions and hence implicitly requires a linear model. In contrast, the EWPF is a fully nonlinear data assimilation method that does not rely on a particular error distribution. The EWPF has been demonstrated to work well in highly nonlinear situations, like in a model solving a barotropic vorticity equation, but it is still unknown how the assimilation performance compares to ensemble Kalman filters in realistic situations. Twin assimilation experiments using a square basin configuration of the NEMO model are performed to assess the filter behavior. The configuration simulates a double gyre, which exhibits significant nonlinearity. The LETKF and EWPF are both implemented in PDAF (Parallel Data Assimilation Framework, http://pdaf.awi.de), which ensures identical experimental conditions for both filters. In the experiments, the EWPF does not succeed to outperform the LETKF. Further, the EWPF requires the explicit simulation of model errors, which represents a particular difficulty in the experiments. The model reacts very sensitively to the model error, which is difficult to tune to obtain a sufficient assimilation performance of the EWPF.