Impact of Scale-Dependent Localization with and without Cross-Band Correlation on Hurricane Track Forecasts in the FV3GFS 4DEnVar System

A uniform spatial localization for all scales is commonly applied in the ensemble-based data assimilation (DA) systems to treat sampling errors. Early studies suggest that the optimal effective localization length for the large scale is broader than that for the small scale. Therefore, the improved performance of the ensemble-based DA systems could be potentially gained by applying different amounts of localization to different ranges of scales. This study explores two variants of the scale-dependent-localization (SDL) method. The first SDL variant assumes that the cross-scale correlation is zero (referred to as the SDL-NoCross). The second variant takes into account the cross-scale correlation (referred to as the SDL-Cross).

The four-dimensional ensemble-variational (4DEnVar) hybrid DA system based on the gridpoint statistical interpolation (GSI) system is being transitioned to the next generation global prediction system FV3GFS at NCEP. These two SDL variants are mathematically demonstrated in an EnVar system using the full B-preconditioning and implemented in the GSI-based 4DEnVar system. In this study, the impacts of these two SDL variants are examined for the FV3GFS global NWP with an emphasis on hurricane track forecasts. To apply the SDL method, the full field in the background ensemble forecasts is filtered into two scales to distinguish the synoptic and storm scales. The results from a single observation experiment using a hurricane example show that these two SDL variants produce the analysis increments that are distinguishable from each other and consistent with their own theoretical expectation. SDL-Cross shows more realistic increments than SDL-NoCross. A cycled DA experiment on Hurricane Irma (2017) is ongoing. The impacts of SDL-NoCross versus SDL-Cross on hurricane track forecasts will also be presented at the conference.