Dealing with observation error correlations and non-Gaussianity in all-sky assimilation
Observation error modelling is an important part of assimilating all-sky satellite radiances. Error variances get much larger in cloudy and precipitating areas because of the error of representation. There are also interchannel and spatial correlations, but these have been ignored in all-sky assimilation until now. New error models have been derived that can change both correlations and variances as a function of cloud amount. Further, they can now be used alongside variational quality control which is also critical for all-sky assimilation. The new models have been applied to hyperspectral infrared radiances, giving the first good all-sky infrared results in a full-featured global forecasting system. They have also been applied to the assimilation of all-sky microwave radiances with significant improvements to forecast scores. In both cases it has been necessary to retune the error covariance matrices to filter out signals that are difficult for data assimilation to handle, such as trace inter-channel biases and gravity waves. The work has also re-examined the role of variational quality control as a way of dealing with the non-Gaussianity of cloud and precipitation errors. The new developments are likely to be applicable for many different observation types.
Dr. Alan Geer ECMWF