Ensemble-4DVAR for the NCEP hybrid GSI-EnKF data assimilation system and observation impact study with the hybrid system

Xuguang Wang
University of Oklahoma


A Four-dimensional ensemble-variational data assimilation capability is developed for the current 3DVAR-based GSI-EnKF hybrid data assimilation system. Like the classic TL/ADJ 4DVAR, four dimensional analyses are obtained with the Ensemble-4DVAR by fitting observations spanning the assimilation window. Temporal evolution of the error covariance within the assimilation window is realized through the use of ensemble perturbations. Forecasts initialized by the analyses generated by the Ensemble-4DVAR were compared with the 3DVAR-based GSI-EnKF hybrid for both summer and winter periods in 2010 assimilating all operational conventional and satellite observations. The presentation will focus on the summer period test. Various verification including global forecasts and hurricane track forecasts showed that the Ensemble-4DVAR improved upon the 3DVAR based hybrid. Strong balance constraint applied on the ensemble covariance was found to degrade the hurricane track forecasts but benefit the general global forecasts. Various extensions to Ensemble-4DVAR is added and being tested for the 2011 hurricane season and for the preparation of the coming hurricane season.

Recent results studying the impact of observations using the hybrid GSI-EnKF hybrid will also be presented. The impacts of observations assimilated using the classic GSI and the hybrid GSI-EnKF were compared for a winter month in 2010 through Observation System Experiment (OSE). Data denial experiments were conducted to access impacts of observations of interest including both conventional observations and satellite observations such as AMSU radiances. It was found that forecasts by the hybrid GSI-EnKF was better than GSI in both the control and data denial experiments. In some verification, the hybrid assimilating less data was better than the GSI assimilating all observations. It was also found the magnitude and distribution of observation impacts depended not only on the types of observations but also whether GSI or hybrid GSI-EnKF data assimilation methods are used. The relative impact between rawinsonde and AMSU also depended on whether GSI or hybrid GSI-EnKF data assimilation was used and also the types of verifications. In addition to the OSE, the ensemble based observation impact metric has been applied to estimate the impact of observations assimilated by the hybrid GSI-EnKF. Initial results show that ensemble based observation impact metric can provide good estimate of the impact and the quality of the estimate can depend on the types of observations and the forecasts of interest. Progress and recent results on unifying and testing the GSI-EnKF hybrid for regional modeling systems will be updated in the seminar as time permits.