Application of a Scale-Separation Verification Technique to Regional Forecast Models

Abstract

A scale-separation technique based on two-dimensional Fourier decomposition is applied to the comparison and verification of analyses and forecasts produced by regional numerical weather prediction systems. A major emphasis of this study is the verification of secondary or derived parameters in addition to the evaluation of primary model variables. Two prediction models are used to illustrate the technique for a variety of forecast fields separated into three separate wavenumber bands. Three different sets of analyses, one from each model system and an independent set, are used for both analysis intercomparison and model verification. The comparison of the analyses is essential to establishing the level of uncertainty for each variable as a function of scale. The synoptic-scale database used to produce the analyses for this study does not allow the verification of scales 800 km or less, no matter how fine the resolution of the model. Examining the spectra of difference fields with time allows one to study the evolution of model error (or differences between two models) as a function of wavenumber. In some instances where traditional statistical measures of skill indicated good agreement between two forecasts, spectral scale selection of the difference fields shows that the spatial distribution of the errors was quite different, pointing to different error-growth characteristics of the models. The technique allows one to partially separate phase and amplitude errors and, hence, barotropic-versus baroclinic-type error structure. It was found, as expected, that forecast skill decreases more rapidly with time for smaller scales, but this is not true for all parameters examined. The presence of lateral boundary conditions strongly influences the evaluation of skill in a regional model for the primary variables, but not as much for some secondary variables. Verification of secondary variables nearly always indicates significant errors in the forecast before serious problems in the primary variables are detected.