Erin Jones

Weather and Climate Systems

Erin Jones

Flow-Dependent Vertical Localization in GFS Hybrid 4DEnVar for Improvement of Global and Tropical Cyclone Track Numerical Prediction

April 24th, 2024

3:30 pm

NWC 1350

Abstract: In ensemble-based data assimilation for numerical weather prediction, limitations in ensemble size due to computational cost restrictions introduce spurious background forecast error correlations.  Covariance localization, which typically applies a modulation function to reduce correlations distant from observation locations, is a cost-effective way to limit this spurious effect. Ideal covariance localization dampens spurious correlations while retaining correlations with dynamic-based significance. Therefore, localization based on the generic spatial length scales of the atmosphere can aid in discerning between correlations with or without a dynamic connection. Currently, though, most commonly used data assimilation systems utilize a fixed localization value. Several methods have been proposed to improve horizontal covariance localization to reflect the structure or the multiscale nature of the underlying flow.  However, there is limited study on improving vertical covaria
nce localization.

This presentation introduces a new vertical flow-dependent localization (vFDL) scheme to be used within the existing GFS hybrid 4DEnVar for global forecasts.  This new approach uses the percentage of ensemble variance explained by the leading mode at each grid point as the predictor of the vertical localization scales.  It is found that this method is able to identify areas of increased or reduced vertical localization consistent with the underlying dynamic structure of the flow. Global forecasts using vFDL show significant improvement over forecasts using fixed localization at 4–5-day lead times. These improvements are most notable at jet levels, which typically exhibit larger correlation length scales than other tropospheric levels. Additionally, results indicate that vFDL has the potential to improve tropical cyclone track prediction for up to 5 days in lead time compared to an experiment using constant localization.  Diagnostics suggest that vFDL allows for more accurate correl
ations to be used for data assimilation throughout the depth of the tropical cyclone and its surroundings, leading to improvements in the large-scale environmental flow.