Modelling background error correlations with spatial deformations: a case study

A long-term goal in variational data assimilation is to improve the anisotropy of background error correlations. One way to achieve anisotropic correlations is to introduce spatial deformations. This deformation can be specified a priori for instance by using the geostrophic transform (GT) as introduced by Desroziers (1997). The deformation can also be estimated from a purely statistical point of view (Michel, 2013a). The aim of this study is to evaluate the performance of such spatial deformation techniques for the use of background error modelling. A large ensemble of variational assimilations with perturbed observations is set up on a case study with the global ARPEGE model. An anisotropy index and a length scale diagnostic are defined to compare objectively the effectiveness of the deformations. This effectiveness is measured as the ability of the inverse spatial deformations to make the correlations more isotropic or more homogeneous. The results are shown to depend on the vertical level and on the variable. Generally, the statistical deformation is able to reduce the anisotropy while the GT is giving much smaller improvements that are, in this case study, confined to the frontal area of an extratropical cyclone.