Abstract/Summary

Data assimilation is an important component of numerical weather prediction (NWP); it is used to estimate realistic initial conditions. In recent years, hybrid data assimilation, which combines the traditional climatological representation of background error covariances with an ensemble representation, has gained significant traction. This thesis develops and applies hybrid ensemble-variational data assimilation to a convective-scale tropical simplified model (ABC-DA) and an NWP system over the western Maritime Continent (SINGV-DA) to explore its benefits and how it can be better designed for the tropics. Firstly, the hybrid ensemble-variational data assimilation approach (hybrid-En3DVar) was implemented in ABC-DA. Generally, hybrid-En3DVar outperformed the traditional 3DVar data assimilation approaches. The improvements were sensitive to ensemble size, and were less pronounced when the ensemble was very small. The results also highlighted how the sub-optimal background error covariance model in 3DVar, which uses geostrophic balance as a balance constraint, led to erroneous analysis increments in meridional wind and negatively impacted the forecasts. Secondly, the hybrid ensemble-variational data assimilation approach was implemented in SINGV-DA. Without proper tuning, the initial hybrid-En3DVar setup had a relatively neutral impact compared to 3DVar. However, tuning the weightings for hybrid-En3DVar and time-shifting of ensemble perturbations were vital for improving precipitation forecasts and forecast fits to radiosonde humidity and wind compared to 3DVar. The results also highlighted how the autocovariance structures varied between variables, including the presence of robust cross-correlation structures between the moisture and temperature-related variables in the ensemble-derived background error matrix, which could have physical significance over the Maritime Continent. Thirdly, the localisation aspect of En3DVar was modified to allow for variable-dependent localisation (prescribing different localisation length-scales for different variables) and selective multivariate localisation (knocking-out by localisation certain multi-variate error covariances) in ABC-DA. This was to address two limitations of traditional ensemble-variational data assimilation approaches which were pertinent over the tropics. Using selective multivariate localisation was beneficial, particularly when covariances associated with hydrostatic balance were retained and when zonal wind errors were de-coupled from the mass errors in the cross-covariances. The results also showed that variable-dependent localisation could be beneficial if the localisation length-scales were well-tuned for each variable. Both these enhancements within a pure EnVar framework reduced the forecast root-mean-square errors by about 3-4% for zonal wind and mass variables. These benefits may also apply to hybrid-En3DVar.