Real-time multianalysis-multimodel superensemble forecasts of precipitation using TRMM and SSM/I products

This paper addresses real-time precipitation forecasts from a multianalysis -multimodel superensemble. The methodology for the construction of the supe rensemble forecasts follows previous recent publications on this topic. Thi s study includes forecasts from multimodels of a number of global operation al centers. A multianalysis component based on the Florida State University (FSU) global spectral model that utilizes TRMM and SSM/I datasets and a nu mber of rain-rate algorithms is also included. The difference in the analys is arises from the use of these rain rates within physical initialization t hat produces distinct differences among these components in the divergence, heating, moisture, and rain-rate descriptions. A total of 11 models, of wh ich 5 represent global operational models and 6 represent multianalysis for ecasts from the FSU model initialized by different rain-rate algorithms, ar e included in the multianalysis-multimodel system studied here. In this pap er, "multimodel'' refers to different models whose forecasts are being assi milated for the construction of the superensemble. "Multianalysis'' refers to different initial analysis contributing to forecasts from the same model . The term superensemble is being used here to denote the bias-corrected fo recasts based on the products derived from the multimodel and the multianal ysis. The training period is covered by nearly 120 forecast experiments pri or to 1 January 2000 for each of the multimodels. These are all 3-day forec asts. The statistical bias of the models is determined from multiple linear regression of these forecasts against a "best'' rainfall analysis field th at is based on TRMM and SSM/I datasets and using the rain-rate algorithms r ecently developed at NASA Goddard Space Flight Center. This paper discusses the results of real-time rainfall forecasts based on this system. The main results of this study are that the multianalysis-multimodel superensemble has a much higher skill than the participating member models. The skill of this system is higher than those of the ensemble mean that assigns a weight of 1.0 to all including the poorer models and the ensemble mean of bias-re moved individual models. The selective weights for the entire multianalysis -multimodel superensemble forecast system make it superior to individual mo dels and the above mean representations. The skill of precipitation forecas ts is addressed in several ways. The skill of the superensemble-based rain rates is shown to be higher than the following: (a) individual model's skil ls with and without physical initialization, (b) skill of the ensemble mean , and (c) skill of the ensemble mean of individually bias-removed models. The equitable-threat scores at many thresholds of rain are also examined fo r the various models and noted that for days 1-3 of forecasts, the superens emble-based forecasts do have the highest skills. The training phase is a m ajor component of the superensemble. Issues on optimizing the number of tra ining days is addressed by examining training with days of high forecast sk ill versus training with low forecast skill, and training with the best ava ilable rain-rate datasets versus those from poor representations of rain. F inally the usefulness of superensemble forecasts of rain for providing poss ible guidance for flood events such as the one over Mozambique during Febru ary 2000 is shown.