A GSI-based hybrid ensemble-variational data assimilation system and its comparison with GSI and ensemble Kalman filter

Data assimilation has been dominated by two parallel methods, the variational method (VAR) and the ensemble Kalman filter (EnKF). Recent studies suggested a hybrid ensemble-variational (hereafter, hybrid) system can take advantages of both methods and therefore provide better forecasts than either method alone. A hybrid data assimilation system was recently developed based on the NCEP operational Gridpoint Statistical Interpolation (GSI) data assimilation system and the experimental ensemble Kalman filter (EnKF). In this system, the extended control variable method was adopted to incorporate the flow-dependent ensemble covariance in the variational framework and the ensemble was generated by the ensemble Kalman filter. The current hybrid system has two tunable parameters, one is the localization scale for the ensemble covariance and the other is the weighting factor that determines how much we trust the static and ensemble covariance. The system is also built to accommodate the analysis and ensemble forecasts at two different resolutions, and the EnKF can be one way or two way coupled with the GSI based hybrid. The hybrid system was first built based on the 3DVAR version of the GSI, and tested for the Global Forecast System (GFS) model. Our initial test shows that forecasts initialized with the hybrid GSI (3DVAR)-EnKF DA is more accurate than that of GSI and comparable to EnKF. More detailed comparison among the three is being conducted for both a retrospective winter month and a summer month. We will present the optimal localization and weighting factors in the hybrid, the impact of one way and two way coupling between the hybrid and EnKF, and the relative performance of the hybrid and GSI and EnKF. Controlled experiments are also being conducted to further understand the difference between a hybrid and a pure EnKF to isolate the impact of the algorithmic differences (e.g., covariance localization and 3D vs. 4D flow dependent ensemble covariance) between the two in their performance. The above will be evaluated with the regular root mean square error and anomaly correlation of wind and temperature forecasts in both warm and cold seasons, and the skill of hurricane forecasts in warm season to understand the performance difference of different data assimilation schemes at different scenario and with different metric.