Many attempts to assimilate precipitation observations in numerical models have been made, but they have resulted in little or no forecast improvement at the end of the precipitation assimilation. This is due to the nonlinearity of the model precipitation parameterization, the non-Gaussianity of precipitation variables, and the large and unknown model and observation errors.

In this study, we investigate the assimilation of global large-scale satellite precipitation using the local ensemble transform Kalman filter (LETKF). The LETKF does not require linearization of the model, and it can improve all model variables by giving higher weights in the analysis to ensemble members with better precipitation, so that the model will "remember" the assimilation changes during the forecasts. Gaussian transformations of precipitation are applied to both model background precipitation and observed precipitation, which not only makes the error distributions more Gaussian, but also removes the amplitude-dependent biases between the model and the observations. In addition, several quality control criteria are designed to reject precipitation observations that are not useful for the assimilation.

Our ideas are tested in both an idealized system and a realistic system. In the former, observing system simulation experiments (OSSEs) are conducted with a simplified general circulation model; in the latter, the TRMM Multisatellite Precipitation Analysis (TMPA) data are assimilated into a low-resolution version of the NCEP Global Forecasting System (GFS). Positive results are obtained in both systems, showing that both the analyses and the 5-day forecasts are improved by the effective assimilation of precipitation. We also demonstrate how to use the ensemble forecast sensitivity to observations (EFSO) to analyze the effectiveness of precipitation assimilation and provide guidance for determining appropriate quality control. These results are very promising for the direct assimilation of satellite precipitation data in numerical weather prediction models, especially with the forthcoming Global Precipitation Measurement (GPM) sensors.