Model-space localization in serial ensemble filters

Ensemble-based data assimilation systems typically use covariance localization to dampen spurious correlations associated with sampling error while increasing the rank of the covariance estimate. Variational methods use model-space localization, in which localization is applied to ensemble estimates of covariances between model variables and is based on distances between those variables, while ensemble filters apply observation-space localization to estimates of model-observation covariances, based on distances between model variables and observations. It has been shown that for nonlocal observations, such as satellite radiances, model-space localization can be superior. This paper demonstrates a new method for performing model-space localization in serial ensemble filters using the linearized observation operators (or Jacobians). Results of radiance-only assimilation in a global forecast system show the benefit of using model-space localization relative to observation-space localization. The serial ensemble square root filter with vertical model-space localization gives results similar to those of the Ensemble Variational system (without outer loops or extra balance constraints) while increasing the runtime compared to the filter with observation-space localization by a factor between 2 and 8, depending on how sparse the Jacobian matrices are. The results are also similar to another approach to model-space localization in ensemble filters: ensemble Kalman filter with modulated ensembles.