This talk is on ensemble Kalman filter (EnKF) experiments with a Primitive-Equations model, a dissertation directed by Prof. Eugenia Kalnay and defended on June 7 at the University of Maryland, College Park. The ultimate goal is to develop a path towards an operational ensemble Kalman filtering (EnKF) system. Several approaches to EnKF for atmospheric systems have been proposed but not compared, and the sensitivity of EnKF to the imperfections of forecast models is unclear. This research explores two basic questions: 1. What are the relative advantages and disadvantages of the two most promising EnKF methods? 2. How large are the effects of model errors on data assimilation and does model bias correction work?
We apply two EnKF methods, serial EnSRF (serial ensemble square root filtering, Whitaker and Hamill 2002) and LEKF (local ensemble Kalman filtering, Ott et al. 2002; 2004), as well as 3DVAR to the SPEEDY Primitive-Equations global model (Molteni 2003). The SPEEDY model is a fast but relatively realistic model allowing a comparison of methods addressing the first question. Our results show that in a perfect model scenario the EnKF outperforms 3DVAR. Surprisingly, the 2-day forecast "errors of the day" are very similar to the analysis errors, but not similar among different methods. In ensemble low-dimensional regions, however, the errors show some similarity. Overall, our results suggest serial EnSRF outperforms LEKF, but their difference is substantially reduced when we localize the error covariance or increase the ensemble size. Since LEKF is much more efficient with parallel computers and many observations, it would be the only feasible choice in operations.
Then, we remove the perfect model assumption and investigate the second question, using the NCEP/NCAR reanalysis as the nature run.
The advantage of EnKF over 3DVAR is greatly reduced. When we apply the model bias estimation proposed by Dee and da Silva (1998), we find that the full dimensional model bias estimation fails. However, if instead we assume the bias is low dimensional, we obtain a substantial improvement.