Sensitivity of a Global Forecast Model to Initializations with Reanalysis Datasets

Abstract

A global research model is initialized with reanalysis datasets obtained from NCEP and ECMWF. The globally averaged accuracy of the resulting 120-h predictions varies little between the different initializations, but a perceptible difference arises in the mid- to high latitudes of the Southern Hemisphere, where ECMWF initialized forecasts have somewhat greater skill. Most of this benefit is explained by differences of the longer-wave components (wavenumbers 0?15) of the initial data. This motivates further diagnoses of globally computed sensitivity measures to initial data changes. Approximately 67% of the 120-h forecast difference produced by changing initial data from ECMWF to NCEP reanalyses is due to initial changes only in wavenumbers 0?15, and more than 85% of this difference is produced by initial changes in wavenumbers 0?20. The result implies downscale uncertainty growth and contradicts several recent predictability investigations based upon singular vector analyses, which emphasize upscale uncertainty growth. The results do not imply that singular vector analyses are in error. They only suggest that large-scale errors of the initial state may play a more prominent role than suggested in some singular vector analyses. Downscale uncertainty evolution may be due to greater analysis uncertainty at large spatial scales than considered in prior recent studies emphasizing upscale predictability loss.