Statistical Downscaling of 2D Reanalysis Fields with the Analog Ensemble

Reanalyses provide a valuable source of information on meteorological parameters as retrospective time series. In order to exploit these data, a statistical downscaling approach can be applied. Such an approach acts as a correction of systematic errors in the reanalysis with respect to the parameter (predictand) in consideration. Reanalysis intercomparison has shown the value of using very high model resolution combined with the assimilation of radar data with respect to the representation of precipitation in the regional reanalysis COSMO-REA2 at 2km horizontal resolution (Wahl et al. 2017). However, due to the availability of radar data, COSMO-REA2 only covers the period from 2007-2013. In our downscaling approach, we use the analog ensemble technique to generate synthetic time series of 2D precipitation fields for earlier time periods. Here, we use the 20-year (1995-2014) data set COSMO-REA6 at 6km resolution as predictor. The probabilistic estimates for the precipitation are created by finding analogs in the existing COSMO-REA2 data set for each time step in the reconstruction period (1995-2006). The analogs are determined as the N best matches in terms of a metric calculated using COSMO-REA6. These analog ensembles are then validated against independent observations over central Europe.