Benjamin MENETRIER, CNRM-GAME / Météo-France

Optimized localization and hybridization to filter ensemble-based covariances



Localization and hybridization are two methods used in ensemble data assimilation to improve the accuracy of sample covariances. Following the work of Ménétrier et al. (2015a,b) about optimal localization diagnostics, Ménétrier and Auligné (2015b) has provided the theoretical background and a practical method to optimize both localization and hybridization coefficients simultaneously. The theory combines the statistics of sample centered moments within a linear filtering framework, and does not assume a Gaussian distribution for the ensemble. The practical implementation uses data from the ensemble only and is affordable for high-dimensional systems. It takes both ensemble size and sample covariance error structures into account, in order to reduce the sampling noise while preserving the signal of interest. Theoretical and experimental evidence shows that if optimal weights are used, localized-hybridized sample covariances are always more accurate than their localized-only counterparts. An open-source code has been made available by authors to diagnose localization functions and hybridization weights. Its generic core can handle any kind of grid, structured or unstructured. It has been successfully tested on several atmospheric and oceanic models.

File available here