Ensemble variational assimilation for the representation of background error covariances in a high-latitude regional model

Statistical objective analysis requires the explicit specification of the observation and background error covariances. This paper deals with the estimation of the latter within a high-latitude regional model. Four different approaches have been adopted to simulate the error evolution in the analysis and forecast steps of the model: i) the largely-adopted NMC method, applied to both a winter and a summer season data set, ii) global ensemble analyses projected forward to the 6-hour forecast range by the limited-area model itself, iii) limited-area ensemble variational assimilation with unperturbed lateral boundary conditions (LBCs) and iv) limited-area ensemble variational assimilation with perturbed lateral boundary conditions. The structure of the four background error covariance matrices was extensively compared. It turned out that the NMC-derived standard deviations had stronger amplitude at the large scales, especially in the lower troposphere, compared to those derived using the ensemble technique, and much broader horizontal and large-scale vertical correlations. The contribution of lateral boundary perturbations is shown to be significant. Moreover, neglecting these LBCs perturbations contributes to unrealistic features, such as artificial small variances near the boundaries, which tend to spuriously propagate towards the inner area. Furthermore, humidity variances and associated cross-covariances tend to be weaker in the ensemble assimilation experiments, in particular when lateral boundary conditions are not perturbed. A one month assimilation period allowed us to evaluate the impact of the different background error statistics on the forecasts: wind, geopotential and partly temperature are clearly benefiting from the use of ensemble-based background error covariances; humidity skill scores are noticeably improved when background error covariances from limited-area ensemble assimilation were used. The seasonality of the background error structures was also investigated, and it was found non-negligible for the performance of the data assimilation and forecast systems.