ESTIMATING OBSERVATION ERROR STATISTICS FOR ATMOSPHERIC DATA ASSIMILATION

All advanced data assimilation systems require reasonable estimates of the observation error statistics for each observation system used in the assimilation. Without such estimates, no assimilation system can extract all the available information from the observations. Observation error has two components - instrument error and the error of representativeness. This latter error can be associated with the error of the forward interpolation operator. This connection is examined in a simple model context and the spatial and temporal characteristics (and correlation with the signal) of the forward interpolation error are explored. Observation error statistics are determined for radiosonde networks using conventional innovation covariance techniques, which are applicable because it can be assumed that radiosonde observation errors are neither horizontally correlated nor correlated with forecast errors. For satellite observation systems, these assumptions are not always appropriate, and conventional innovation techniques cannot be used. A generalized innovation crosscovariance method for determining observation error statistics for satellite systems is proposed. In principle, this technique can determine spatial correlations and correlations between forecast and observation error. It is successfully tested in a simple model context.