"Understanding the differences between EnVar and LETKF solvers in an operational NWP setting"

The International Symposium on Data Assimilation Online (ISDA-Online) 8 Jan,2021 "Data Assimilation Methodology"
International Symposium on Data Assimilation - Online

In the next upgrade, the NOAA operational global hybrid ensemble-variational data assimilation system will implement an LETKF solver with model-space vertical localization to update ensemble perturbations. Previous work has shown that the inferior performance of the serial EnSRF and LETKF solvers compared to the (nonhybrid) EnVar solver in the NOAA system was primarily due to the of observationspace localization in the EnKF when assimilating radiances. Now that model-space localization has been implemented, the LETKF solver seems to perform slightly better than EnVar. In this talk, a hierarchy of simpler models is utilized to understand the reason for this. The results show that observation-error (R) localization (used in the LETKF) outperforms covariance (B) horizontal localization (used in EnVar) under certain conditions. In particular, when the horizontal scale of the Kalman Gain is narrower than the horizontal scale of the background-error covariance, R-localization performs better since it acts directly to localize the gain matrix. This tends to occur in the simple models studied when there are dense and/or accurate obs (the 'strong assimilation' limit) and the background-error covariance has heavy tails (the covariance is closer to exponential than it is to Gaussian). Both of these conditions are present in the operational NWP setting.