Effects of Observation Errors on the Statistics for Ensemble Spread and Reliability

Abstract

The effects of observation errors on rank histograms and reliability diagrams are investigated using a perfect model approach. The three-variable Lorenz-63 model was used to simulate an idealized ensemble prediction system (EPS) with 50 perturbed ensemble members and one control forecast. Observation errors at verification time were introduced by adding normally distributed noise to the true state at verification time. Besides these simulations, a theoretical analysis was also performed. One of the major findings was that rank histograms are very sensitive to the presence of observation errors, leading to overpopulated upper- and lowermost ranks. This sensitivity was shown to grow for larger ensemble sizes. Reliability diagrams were far less sensitive in this respect. The resulting u-shaped rank histograms can easily be misinterpreted as indicating too little spread in the ensemble prediction system. To account for this effect when real observations are used to assess an ensemble prediction system, normally distributed noise based on the verifying observation error can be added to the ensemble members before the statistics are calculated. The method has been tested for the ECMWF ensemble forecasts of ocean waves and forecasts of the geopotential at 500 hPa. The EPS waves were compared with buoy observations from the Global Telecommunication System (GTS) for a period of almost 3 yr. When the buoy observations were taken as the true value, more than 25% of the observations appeared in the two extreme ranks for the day 3 forecast range. This number was reduced to less than 10% when observation errors were added to the ensemble members. Ensemble forecasts of the 500-hPa geopotential were verified against the ECMWF analysis. When analysis errors were neglected, the maximum number of outliers was more than 10% for most areas except for Europe, where the analysis errors are relatively smaller. Introducing noise to the ensemble members, based on estimates of analysis errors, reduced the number of outliers, particularly in the short range, where a peak around day 1 more or less vanished.