Ensemble forecasting and data assimilation

Istvan Szunyogh and Eugenia Kalnay

Department of Meteorology University of Maryland


In this presentation, we explain the implementation of the Local Ensemble Kalman Filter (LEKF, Ott et al. 2003 ) on the T62, 28-level version of the full operational NCEP GFS model. We demonstrate that the LEKF scheme is efficient in assimilating a large number of observations, as it is well suited to a massively parallel computing environment. Our experiments, assimilating simulated observations (obtained by perturbing a known true state), show that, with the current version of the code, the assimilation of 1.7 million observations takes about 6 minutes on a 40-processor cluster of 2.8 GHz Xeon processors (a $150,000 computer). Also, assimilating observations at a mere 10% of the model grid points provides analyses that are almost as accurate as those obtained by observing the atmospheric state at all grid points.

We point out that the ensemble based formulation allows for a practical way of unifying the extended Kalman filter and the four-dimensional variational approach, in what we call Four-Dimensional Ensemble Kalman filter (4DEKF, or 4DLEKF in the local setting). In this approach we can take exact observation times into account in a natural way, even if they are different from the assimilation time. Essentially we propagate the observational increments at intermediate time steps using the ensemble of background forecasts.