On errors in meteorological data assimilation

Data assimilation in Numerical Weather Prediction (NWP) optimally blends observations with atmospheric model data in order to obtain the best possible initial state for an atmospheric model prediction. Specification of error characteristics is an important part of data assimilation. This thesis is concerned with representation of background error standard deviations, with handling of observations, and with observation error characteristics. The research includes both the study of basic assimilation problems within the framework of an idealised quasi-geostrophic model and the development of assimilation algorithms for a full scale limited area high resolution forecasting system. It is shown in this thesis that an accurate representation of background error standard deviations is important for the quality of NWP forecasts. In particular the effect of introducing a time-dependency is investigated and a novel approach to relate the flow-dependency of background error standard deviations to an Eady baroclinic instability measure is developed. The Eady based flow-dependent background error representation is demonstrated to have a positive impact on NWP, as compared to horizontally and temporally independent background error statistics. An alternative method, based on on-line error estimation and maximum likelihood theory, is proven to be able to represent the flow-dependency of background error standard deviations and encouraging results are obtained within the quasi-geostrophic model framework. Furthermore, it is shown that a proper observation handling is an important part of data assimilation. The treatment of error characteristics is specifically shown to be of major importance when exploiting the potential benefit of radar radial wind observations within data assimilation