Approaches to convective scale data assimilation

Kurzfassung/Abstract

The applications of data assimilation on convective scales require a numerical model of the atmosphere with single digit horizontal resolution in km and time evolving error covariances. Past studies have shown that ensemble Kalman filter (EnKF) algorithm can be applied to the convective scales since it is capable of handling complex and highly nonlinear processes. However, some challenges for the convective scale applications still remain. These include a need to estimate on convective scale, fields that are nonnegative (such as rain, graupel, snow) and to use data sets such as radar reflectivity or cloud products that have the same property. What underlines these examples are errors that are non-Gaussian in nature causing a problem with the EnKF that uses Gaussian error assumptions to produce the estimates from the previous forecast and the incoming data. Since the proper estimates of hydrometeors are crucial for prediction on convective scales, the question arises whether the EnKF method can be modified to improve these estimates and whether there is a way of optimizing the use of high-resolution data such as radar observations or ModeS wind data (with rapid updates) to initialize numerical weather prediction models. In this talk, we first review the challenges that convective scale data assimilation methods are facing when assimilating radar data. This is done using the non-hydrostatic convection permitting COSMO model (Baldauf et al. 2011), and the local ensemble transform Kalman filter (LETKF, Hunt et al. 2007) as implemented in KENDA (Km-scale Ensemble Data Assimilation) system of German Weather Service (Schraff et al. 2016). Currently KENDA uses latent heat nudging for radar data assimilation (see Schraff et al. 2016), but the transition to localized EnKF for radar data has been tested as well (Bick et al. 2016). We will show that due to the Gaussian assumptions that underline the LETKF algorithm, the analyses of water species will become negative in some grid points of the COSMO model. These values are set to zero after the LETKF analysis step, in order not to give the numerical model unphysical values. The tests done within this setup show that such a procedure introduces a bias in the analysis ensemble with respect to the truth, that increases in time due to the cycled data assimilation. Further, the noise produced by assimilation of radar reflectivity with LETKF is larger than if only conventional data (including the high resolution ModeS data) are assimilated. If the horizontal resolution of numerical model is changed from 2.8 km (COSMO-DE operational resolution) to 1.4 km, the noise even increases further and surpasses the noise levels of experiments that used latent heat nudging for assimilation. In all EnKF experiments performed, assimilation of the radial wind measurements in addition reduces the noise. In order to better understand some of the above challenges, the localized EnKF has been tested in an idealized framework