Data Assimilation: the 1984/85 Revisions of the Ecmwf Mass and Wind Analysis

Abstract A comprehensive evaluation of the ECMWF operational data assimilation system has been made, based largely on statistics of its performance over a long period of time. the evaluation has identified many deficiencies, covering such aspects of the analysis as the optimum interpolation statistics, the data selection algorithms, and the quality control criteria applied within the analysis. The observation and forecast error variances were seriously in error in the stratosphere and in the tropics. A more faithful model of the forecast error spectrum is provided by the Bessel function expansion than by the Gaussian model. the vertical structure of the height error is assumed to be composed of two components: a ‘barotropic’ part related to a horizontally constant mode and a ‘baroclinic’ structure for the horizontally varying modes. the wind field errors also have a large‐scale component. the extratropical vertical correlation remained similar to the earlier correlations, but the tropical height and wind structures changed dramatically. Reasonable rejection limits were defined from the distributions of the observation departures from the first guess, as were observation errors for satellite and aircraft wind reports and Australian pseudo‐observations. A comprehensive and homogeneous data selection algorithm has been devised. However, it has been found necessary to exclude land surface wind reports from the analysis process as the wind departures from a first‐guess field commonly contradict the surface pressure departures due to difficulties in generating a reasonable 10 m first‐guess wind. the omission of surface winds alleviated the data selection problem in data‐dense regions. The representation of the vertical correlations was changed to a continuous representation in the vertical which enabled a multivariate interpolation from data on standard pressure levels directly to model coordinates. the earlier procedure used a multivariate interpolation to constant pressure surfaces in the model grid, followed by a univariate interpolation to model surfaces in the vertical. The results of these improvements, in a series of tests covering in total some 30 days of assimilations, are also described. In summary they yield an analysis and a subsequent first guess which are in closer agreement with observations. the analysis scheme yields more balanced fields in the sense that the subsequent initialization produces more modest changes. the major impact on forecasts, in particular in the short range, is also judged to be positive. These improvements have been incorporated in the ECMWF operational data assimilation system. They have also been incorporated in the reassimilation of the final FGGE level IIb data set at ECMWF.