A Comparative Convective Study Between the Ensemble Kalman Filter and the Local Particle Filter Using the Gridpoint Statistical Interpolation System

Abstract

The localized particle filter (LPF) is a recent advance in ensemble data assimilation for numerical weather prediction. Derived from the original particle filter used for highly nonlinear state variables, the LPF incorporates a localization function to reduce the influence of distant observations. Particle filters are an effective data assimilation method for higher order variables and is widely used in the geoscience community for its ability to maintain nonlinear properties through cycling. The LPF has been proven to be successful in idealized cases. This work seeks to evaluate the LPF for real-data convective-scale weather predictions.

This study compares the performance of storm-scale analyses and predictions generated from the LPF method compared to the classic ensemble Kalman filter (EnKF) which is commonly used in atmospheric data assimilation weather prediction systems. Since the LPF does not contain many of the underlying assumptions that the EnKF does, it is hypothesized that the LPF may be useful for convective-scale data assimilation in generating accurate analyses and forecasts. This research project uses NSSL’s Warn-on-Forecast System (WoFS) to compare the performance of the two data assimilation schemes.

The use of the LPF provides some benefits over the EnKF when producing smaller posterior and prior root-mean-square-errors (RMSE) for non-Gaussian variables, such as reflectivity. More linear variables, such as radial wind, are assimilated at a similar efficacy. Ensemble members of the LPF create more spread of dewpoint temperature than the EnKF within the mid-levels of the atmosphere. Overall, the LPF tends to create a dry bias within the environment, leading to premature decay of the storms.