9:00 AM - 9:15 AM
[MGI24-01] Reduced non-Gaussianity and improved analysis by assimilating every-30-second radar observation with ensemble Kalman filter: a case of idealized deep convection
★Invited Papers
Keywords:Data Assimilation
The non-Gaussian error probability distribution is a major challenge of convective-scale application of the ensemble Kalman filter (EnKF). In the case of radar reflectivity assimilation for rapidly growing convective systems, nonlinear state evolution makes the error distribution strongly non-Gaussian even within 5 minutes. The non-Gaussian error distribution is considered as one of the difficulties in estimating model variables from the observation by weather radars (Fabry and Meunier 2020).
The assimilation of radar observation with the interval shorter than 5 minutes is an interesting approach to tackle this problem. Previous studies assimilated the phased array weather radar (PAWR) data every 30 seconds using the local ensemble transform Kalman filter (LETKF) to capture the rapid growth of localized intense thunderstorms (Miyoshi et al., 2016; Maejima et al., 2017; Honda et al., 2022). Recently, Ruiz et al. (2021) investigated the non-Gaussianity by assimilating the Osaka PAWR data using the 1000-member 4D-LETKF with various assimilation intervals and showed significant impacts on non-Gaussianity and vertical wind analysis.
Since the previous studies investigated real-world cases, it was difficult to accurately verify the analysis. Additionally, it is difficult to distinguish the effect of non-Gaussianity from other factors which may also degrade the analysis and forecast, such as the error in model physics, observation operators, limited observation coverage, and multi-scale background error.
In this study, we perform a series of idealized OSSEs to investigate the impact of assimilating radar observation at higher frequency, focusing on the non-Gaussianity and the analysis accuracy. We carefully design the idealized experiments to exclude other complex factors.
We performed idealized experiments of the assimilation of radar observation in a supercell case initiated by a warm bubble, using 100-member LETKF and synthetic radar reflectivity observation every 30 seconds generated from the nature run. We made the initial ensemble perturbations by shifting the location of the initial warm bubble and adding random band-passed temperature perturbations. The model physics and observation operator, and the background profiles are common between the nature run and the data assimilation experiments, so that we focused on uncertainties only in a convective scale. We compared the analysis fields after 50 minutes of data assimilation cycle of three different cases: namely, every 5 minutes assimilation (5MIN-3D), every 5-minutes 4D assimilation using all observation within past 5 minutes (5MIN-4D), and every 30 second assimilation (30SEC).
Figures 1a-1c show the analysis mean vertical velocity (contours) and its deviation from the nature run (color shades) in x-z cross sections. The 5MIN-3D and 5MIN-4D cases underestimate the updraft near and above the maximum position, whereas the 30SEC case almost perfectly reproduces the true vertical velocity field. Figures 1d-1f show the ensemble spread (red contours) and KL divergence calculated against the Gaussian distribution (color shades) of vertical velocity for each case. The case 5MIN-3D shows the largest spread and KL divergence in vertical velocity. The 5MIN-4D shows smaller values, though they are still significantly larger than that of 30SEC.
We demonstrated that assimilating radar reflectivity every 30 seconds leads to significant reduction of the non-Gaussianity of the ensemble and the improvement of the analysis field, particularly for vertical velocity associated with the strong convection. We also found that its impact on the 30-minute-lead precipitation forecast is limited in these idealized settings without model error and background error in spatial scales larger than mesoscale.
The assimilation of radar observation with the interval shorter than 5 minutes is an interesting approach to tackle this problem. Previous studies assimilated the phased array weather radar (PAWR) data every 30 seconds using the local ensemble transform Kalman filter (LETKF) to capture the rapid growth of localized intense thunderstorms (Miyoshi et al., 2016; Maejima et al., 2017; Honda et al., 2022). Recently, Ruiz et al. (2021) investigated the non-Gaussianity by assimilating the Osaka PAWR data using the 1000-member 4D-LETKF with various assimilation intervals and showed significant impacts on non-Gaussianity and vertical wind analysis.
Since the previous studies investigated real-world cases, it was difficult to accurately verify the analysis. Additionally, it is difficult to distinguish the effect of non-Gaussianity from other factors which may also degrade the analysis and forecast, such as the error in model physics, observation operators, limited observation coverage, and multi-scale background error.
In this study, we perform a series of idealized OSSEs to investigate the impact of assimilating radar observation at higher frequency, focusing on the non-Gaussianity and the analysis accuracy. We carefully design the idealized experiments to exclude other complex factors.
We performed idealized experiments of the assimilation of radar observation in a supercell case initiated by a warm bubble, using 100-member LETKF and synthetic radar reflectivity observation every 30 seconds generated from the nature run. We made the initial ensemble perturbations by shifting the location of the initial warm bubble and adding random band-passed temperature perturbations. The model physics and observation operator, and the background profiles are common between the nature run and the data assimilation experiments, so that we focused on uncertainties only in a convective scale. We compared the analysis fields after 50 minutes of data assimilation cycle of three different cases: namely, every 5 minutes assimilation (5MIN-3D), every 5-minutes 4D assimilation using all observation within past 5 minutes (5MIN-4D), and every 30 second assimilation (30SEC).
Figures 1a-1c show the analysis mean vertical velocity (contours) and its deviation from the nature run (color shades) in x-z cross sections. The 5MIN-3D and 5MIN-4D cases underestimate the updraft near and above the maximum position, whereas the 30SEC case almost perfectly reproduces the true vertical velocity field. Figures 1d-1f show the ensemble spread (red contours) and KL divergence calculated against the Gaussian distribution (color shades) of vertical velocity for each case. The case 5MIN-3D shows the largest spread and KL divergence in vertical velocity. The 5MIN-4D shows smaller values, though they are still significantly larger than that of 30SEC.
We demonstrated that assimilating radar reflectivity every 30 seconds leads to significant reduction of the non-Gaussianity of the ensemble and the improvement of the analysis field, particularly for vertical velocity associated with the strong convection. We also found that its impact on the 30-minute-lead precipitation forecast is limited in these idealized settings without model error and background error in spatial scales larger than mesoscale.