NOAA Earth System Research Lab, University of Colorado/CIRES

EnKF systems typically use 'observation-space' localization for computational reasons. Observation space localization tapers the sample covariance estimate to zero based on the distance between an observation and model state variables. Ensemble-based variational update algorithms use 'model-space' localization, which only depends on the distance between model state variables. Observation (model) space localization applies localization to the sample covariance estimate after (before) the forward operator is applied in the computation of the Kalman gain. Evidence suggests that ensemble-variational systems make better use of radiance observations than EnKF systems. One reason for this could be that observation-space localization can be problematic for radiances, since the vertical distance between a radiance observation and a model state variable is not well defined. To test this hypothesis, we develop a method for applying model-space localization in the vertical in EnKF systems, based upon the 'modulated-ensemble' approach described in a series of papers by Bishop and Hodyss. In this approach, an augmented ensemble is created which includes the effect of vertical localization in model space, and then the EnKF algorithm is applied as usual, but without vertical localization. Results from an experiment that assimilates only radiance observations with the operational NCEP global EnKF system show that using model space localization significantly improves forecast skill. Strategies for optimizing the model-space vertical localization algorithm are discussed, and it is suggested that the computational cost of the model-space approach should not be significantly larger than the observation space approach.

*email: jeffrey.s.whitaker@noaa.gov

*Preference: **Oral **