A JEDI-based 2DVar Snow Depth Analysis for NOAA's Global NWP system

NOAA is upgrading the snow depth analysis for our next global NWP system update, GFSv17. The current state of the art for snow depth analysis in NWP is to use an Optimal Interpolation (OI) to assimilate station snow depth and satellite snow cover. The snow depth OI parameterizes the model error covariances as functions of horizontal and vertical distance. Our first prototype of the GFSv17 snow depth analysis, dubbed the LETKF-OI, approximated the OI by applying an LETKF with an artificially generated ensemble and localization functions that mimicked the OI's assumed model error correlations. Recently, a second prototype has been developed, using 2DVar and applying the JEDI fake levels method to incorporate the OI's vertical correlation structures into the horizontal 2D background error covariances. In tests against independent observations, the 2DVar has similar performance to the (non-JEDI) OI, both of which outperform the LETKF-OI. The 2DVar has then been selected for implementation as the snow depth data assimilation method in GFSv17.

Looking to the future, additional experiments with JEDI have also demonstrated that (true) ensemble-based data assimilation algorithms, such as the LETKF, outperform the above deterministic snow data assimilation methods. This is thought to be due to the dynamic background error variances gained from using an ensemble, since snow depth model errors vary enormously in space and time. Future work will then focus on a potential replacement of the 2DVar snow depth analysis with an ensemble-based analysis. In particular, expanding the current 2DVar into a hybrid EnVar would take advantage of the relative strengths of the ensemble and variational DA methods, while also matching our atmospheric data assimilation method, introducing the option to test strongly coupled land / atmosphere data assimilation.

The above inter-comparison of various different DA algorithms for NOAA's snow depth analysis would not have been achievable without using JEDI. In particular, the use of JEDI greatly reduced the effort involved in switching between DA algorithms, while also ensuring a clean comparison between the algorithms.