The ensemble-based 3DVar (three-dimensional variational technique) method with SVD (singular value decomposition) technique (SVD-En3DVar) is used to assimilate simulated Doppler radial velocity observations. A localization scheme is introduced to the method to reduce spurious error covariance among distant points. The impact of different methods of producing initial perturbation and integration time lengths for forecast samples on the assimilation is emphatically investigated. A new scheme producing initial perturbation samples is proposed. This scheme takes the pseudo-random perturbation fields of temperature and specific humidity as the observation innovation and a 3DVar system is utilized to yield the initial perturbation fields of all variables from the observation innovation. Experiments using the simulated observations by WRF(weather research forecasting) model demonstrate that in the initial perturbation fields produced by using the new scheme the compatibility between different variables is better and the perturbation will not decay quickly in the forecast, so the spin-up time is cut down and the time interval of assimilation cycle can be shortened. The forecasts of temperature, humidity, horizontal wind, and rainfall within 12 hours are improved after assimilating the radar velocity data.