282 IMPACT OF 4DVAR ASSIMILATION OF RAINFALL DATA ON THE SIMULATION OF MCSS VOL. 24

associated with other mei-yu systems. The current networks of conventional observations are unable to provide sufficient information for accurately initializing the Numerical Weather Prediction (NWP) model with sufficient high resolution. Furthermore, the understanding of the physical and dynamical mechanisms of the genesis, evolution and maintenance of the convective precipitation systems within the mei-yu systems is far from complete. Therefore, research efforts aimed at trying to improve the estimates for the initial conditions, thereby allowing reductions in the uncertainties, of the forecasting of the mei-yu weather systems are challenging, yet very valuable. One way to improve these estimates is through the assimilation of rainfall data that are often available from dense surface rain gauge networks. Although the uncertainty in the lateral boundary conditions, as well as the model uncertainties, can also be significant, we neglect them here in order to isolate the uncertainties in the initial conditions. Rainfall observations were first incorporated into numerical weather prediction (NWP) models in the literature through initialization processes. Many initialization schemes, such as the dynamical initialization (Fiorino and Warner, 1981), physical initialization (Krishnamurti et al., 1991, 1993) and cumulus initialization (Donner, 1988) have been developed and notable positive impacts have been reported. In the initialization processes, the rainfall observations are mainly used to modify the thermodynamic properties, such as the temperature and moisture fields, so as to remedy the precipitation spin-up problem and to improve the short-range quantitative precipitation forecasts. All these studies have shown the effectiveness of these initialization schemes. However, initialization is an indirect way of using precipitation data, after all, it cannot guarantee the dynamical consistency among all of the initial fields. Directly using rainfall observations via four-dimensional variational data assimilation (4DVAR) is more desirable, as it can adjust all of the initial fields in a dynamically and physically consistent way. This study focuses on the use of the 4DVAR method and its advantages over other assimilation approaches that have been discussed, (e.g., Zou and Kuo