A comparison of sensitivity of DRP4DVar performance to perturbation samples obtained between the two different methods

The dimensionreduced projectionfour dimensional variational data assimilation (DRP-4DVar) approach utilizes the ensemble of historical forecasts to estimate the background error covariance (BEC) and directly obtains the analysis on the ensemble space. As a result, the quality of ensemble members plays an important role in the performance of the DRP4DVar approach. In this study, the BEC of the Weather Research and Forecast Model (WRF) 3-dimensional variational data assimilation (3DVar) system is employed to produce initial perturbation samples for the DRP-4DVar approach. The historicalforecastbased initial perturbation samples are flowdependent and able to describe the errorgrowing pattern in the atmospheric model and the balanced relationship between different model variables. However, the ensemble spread is not big enough because of the short time interval between adjacent historical samples and the limited numbers of ensemble. The perturbation method of control variables based on the structure characteristics of 3DVar BEC is able to produce initial perturbation samples that have reasonable background error correlations. Moreover, the estimated background error covariance also has good consistence and dynamic and physical balance between variables if the initial perturbation samples experience a development through 6 or 12 h model forward integration. From the aspect of computational expense, the historical forecast results can be directly obtained without any additional computational cost from operational numerical weather forecast centers, while the integration of samples from the 3DVarbased control variable method is timeconsuming but this difficulty can be alleviated through parallel computing. With the analysis of the DRP4DVar based on the new sample as the initial condition (IC), the 9 to 12 h forecasts are better, although the IC is a little worse than that from the DRP-4DVar using the historicalforecastbased sample. Moreover, the simulated precipitation is improved significantly when the new sample is used. As a result, the performance of DRP4DVar approach can be improved further, especially in terms of the accuracy of numerical forecasts and the ability to simulate severe precipitation, when using the perturbarion method of control variables of the 3DVar system to generate the initial perturbation sample. Key words