Andrew Lorenc, Met Office

The benefit of generalised "localisation" of ensemble covariances


Ensemble DA schemes in NWP are limited by computer constraints to an ensemble size VERY much smaller than the order of the background error covariance matrix they are used to estimate. Ensemble Kalman Filter (EnKF) algorithms ameliorate this by localising the problem; usually this involves modifying the observation-space covariances by taking a Schur product with a function which drops off with separation.

Ensemble-variational (EnVar) algorithms use the ensemble-based covariance in model space, which makes available a much wider range of methods to filter the estimated covariances; they can be considered generalised "localisation". They include hybridisation with a climatological estimated covariance, spectral localisation, and scale-dependent spatial localisation. We can also augment the ensemble size by using lagged and time-shifted ensembles. In general the availablity of these methods is one advantage of EnVar over the EnKF.

This talk will demonstate a range of such methods on a simple toy problem, and also present results from a series of experiments with the Met Office's EnVar system, measuring their benefit in practical NWP.

File available here