Initial conditions that accurately represent the atmospheric state at all scales are key for accurate numerical weather prediction down to the convective scales, especially for systems involving multi-scale interactions. In this study, multi-scale data assimilation (DA) capability is developed in the GSI EnKF framework for use with the FV3 limited area model (LAM). The goal is to enable the creation of reliable, larger scale EnKF analysis increments when assimilating coarser-resolution observations that sample the broader-scale atmospheric flows (e.g., rawinsonde), while being able to retain small scale structures when assimilating high-resolution observations such as those of radar at sample precipitation regions. This is achieved by applying a low-pass filter to the highresolution ensemble perturbations to enable the use of large covariance localization radii with smooth ensemble covariances with large-scale observations.
In practice, the multi-scale EnKF DA is realized in a sequential manner. Firstly, 3-km ensemble forecasts are smoothed using a Lanczos low-pass filter to remove small-scale perturbations. Conventional observations are then assimilated in GSI EnKF using relatively large covariance localization radii. The application of the filter reduces noise in covariances at larger distances and improves larger scale analysis. A multiscale RTPS procedure is developed and applied after the assimilation of conventional data to restore the ensemble spread only at large scale, avoiding undesired small-scale increments being introduced. The 3 km ensemble perturbations adjusted by the conventional data are then used in the next step assimilating radar data, using covariance localization radii typical of radar DA on convective-scale grids. The multi-scale DA capabilities implemented within GSI EnKF coupled with FV3 LAM model is applied to a Central Plains tornado-spawning storm case in a half-day hourly cycled manner and the results are compared with those using the regular (i.e., single-scale) EnKF method. The detailed evaluation will be presented in the seminar.
