Distribution-respecting Model-space Vertical Localization for the Non-Variational Ensemble Data Assimilation of nonlocal Observations: Formulation and Experiments with Clear-sky Infrared Radiances

Abstract

Operational meteorological ensemble data assimilation (EnsDA) systems frequently assimilate satellite radiance observations. These observations are non-local because they represent spatially-integrated atmospheric information. For ensemble Kalman filters (EnKFs), model-space localization has been hypothesized to be superior to observation-space localization for assimilating nonlocal observations. This study aims to (i) explore the pitfalls of existing model-space localization schemes, (ii) formulate a new model-space localization scheme for both EnKFs and non-parametric EnsDA that avoids those pitfalls, and (iii) demonstrate the impacts of our new scheme for assimilating clear-sky geostationary infrared radiance observations. Our scheme flexibly handles both Gaussian and non-Gaussian prior distributions. Furthermore, our scheme extends model-space localization from covariance-based linear relationships to nonlinear relationships. We then performed non-cycled offline EnsDA observing system simulation experiments (OSSEs) using Navy Global Environmental Model prior ensembles to investigate our proposed scheme. Results indicate that our scheme is potentially superior to observation-space localization for the EnsDA of those infrared observations. This superiority is observed for both EnKF EnsDA setups and partially non-parametric EnsDA setups. The results also suggest that our scheme is more suitable for non-parametric EnsDA than the frequently-used ensemble modulation scheme for model-space localization. Finally, OSSEs with non-Gaussian configurations of our scheme suggest that the prior marginal distributions of specific humidity may be well-modeled by rectified Gaussian distributions. Our scheme potentially advances the assimilation of operationally-relevant nonlocal observations via non-variational EnsDA methods.