The mesoscale initialization adjustments did not consistently improve skill scores by a large amount, although their impact for individual cases was occasionally large. In general, the only statistically significant improvements occurred when radar echo was used to eliminate dry layers, but these impacts were restricted to amounts of 6.35 mm or less in 6 hours. Qualitatively, changes in predicted rainfall fields were relatively small for all initialization adjustments, but were much larger for changes in the convective scheme used. Some impacts showed a dependence on the larger scale synoptic forcing. For instance, in strongly-forced cases, the BMJ runs performed better.
Evaluations of ensembles made up of the variants showed that skill scores were higher for the ensembles than any single deterministic run, but improvements were modest. Runs using the same convective scheme clustered together, with the observed amount often falling between the BMJ predictions and the KF predictions. This suggests better ensemble guidance will require the use of multiple models or additional convective schemes. In other tests using the output of various ensemble members in a more deterministic sense, it was found that the probability of receiving rainfall above a certain threshold is indeed higher when multiple model members forecast it there. In addition, the probability of receiving rainfall exceeding a given threshold is higher in those areas where a model forecasts rainfall exceeding an even heavier threshold.
Finally, we have also investigated rainfall prediction using a neural network system. Training the network on 32 weather parameters derived from rawinsonde data and operational Eta model output over three years for Omaha, Nebraska and Davenport, Iowa, we found that the network outperforms substantially both the 10 km Eta model and operational Eta model for 24 hour QPF and probability of precipitation forecasts.
Supplementary URL: