Eigenvector-spatial localisation

Abstract

We present a new multiscale covariance localisation method for ensemble data assimilation that is based on the estimation of eigenvectors and subsequent projections, together with traditional spatial localisation applied with a range of localisation lengths. In short, we estimate the leading, large-scale eigenvectors from the sample covariance matrix obtained by spatially smoothing the ensemble (treating small scales as noise) and then localise the resulting sample covariances with a large length scale. After removing the projection of each ensemble member onto the leading eigenvectors, the process may be repeated using less smoothing and tighter localizations or, in a final step, using the resulting, residual ensemble and tight localisation to represent covariances in the remaining subspace. We illustrate the use of the new multiscale localisation method in simple numerical examples and in cycling data assimilation experiments with the Lorenz Model III. We also compare the proposed new method to existing multiscale localisation and to single-scale localisation.

Keywords

• localisation
• multiscale
• data assimilation
• kalman filter
• ensemble