Mitigating Sampling Error and Gaussian Approximation Bias in Regional UFS Data Assimilation

In numerical weather prediction (NWP) models, handling sampling errors arising from small ensemble sizes, model errors, and nonlinear model dynamics is a substantial challenge. Increasing the ensemble size brings a promising prospect for improving prediction accuracy; however, it often encounters limitations, particularly when it comes to the computational resources needed for handling large-scale, high-resolution datasets. The current study introduces a method to address the issue of sampling error in convective-permitting data assimilation by enhancing the flow-dependent background error covariances from the filter update with perturbations sourced from a different model. This approach uses the advantages of a hybrid background error covariance matrix, integrating a more comprehensive representation of system behavior. Specifically, we supplement ensemble perturbations from the Global Data Assimilation System (GDAS), and the perturbations are centered on the ensemble mean of the NOAA Hurricane Analysis and Forecast System (HAFS) at each analysis period Additionally, we conduct a comparative analysis of the Ensemble Kalman Filter (EnKF), localized particle filter (LPF), hybrid LPF-EnKF, and EnKF/LPF combined with En3DVar using the augmented perturbation strategy. Experiments that rely on 80 augmented perturbations from GDAS for updating 40-member ensembles with the EnKF and LPF are found to produce substantial improvements over benchmark experiments. Likewise, a hybrid LPF-Var method, which blends the strengths of parametric and non-parametric data assimilation, consistently demonstrates superior performance to the current EnKF/EnVar strategy. The new approaches are evaluated over multi-week cycling data assimilation experiments focusing on Hurricanes Laura and Marco in August 2020.