Investigation of NCEP GFS Model Forecast Skill "Dropout" Characteristics using the EBI Index

The NCEP Global Forecast System (GFS) model suffers from occasional forecast skill busts or �dropouts� from the treatment of quality control (QC) and bias correction of conventional and non-conventional observations in the assimilation system along with possible model errors. In a companion paper we have conducted several sensitivity experiments assimilating the ECMWF gridded initial conditions (IC) as pseudo-observations (ECM runs), and using them as sole input into the NCEP Gridpoint Statistical Interpolation (GSI) analysis. These ECM 5-day forecasts alleviate several cases of Northern Hemisphere (NH) and Southern Hemisphere (SH) dropouts. Synoptically the IC errors derive from areas that are dynamically active. ECM runs provide a way to make controlled experiments to study the interaction of observation types with dynamical regions in a complex operational assimilation system. The goal is to create detection algorithms by identifying sensitive regions from which the GSI analysis errors grow disproportionately with the potential for causing forecast skill dropouts.

One approach to investigating dynamically active regions is to use a relationship between the growth of mid-latitude baroclinic disturbances and model forecast errors. A suitable measure of the baroclinicity or the maximum growth rate of the most unstable mode is provided by the Eady Baroclinic Instability (EBI) index. Taking account of vertical structure resolution and spin-up in the analysis the EBI index is applied to NH and SH GFS dropouts. The results are compared with the corresponding ECM and ECMWF forecasts. Overall, the GFS contains greater baroclinicity in the NH and SH mid-latitude regions compared to ECM and ECMWF 1-5 day forecasts. The EBI also shows some noise at less than 1-day forecasts when using an analysis as a background guess instead of time filtered previous forecasts as the background. Dynamically active areas according to the EBI index intersecting with areas where QC problems in observation types are found, are tested to see their relevance to dropouts. These regions compare reasonably with the sensitive regions shown by adjoint sensitivity results found at other national centers. Further diagnostics of the forecast errors between dropouts and nondropouts from GFS, ECM, and ECMWF models are shown.