coverage and vertical extent for initializing prediction models. Advances in our predictive capabilities will require a better matching of the observational capabilities with chemical weather prediction needs (Carmichael et al., 2008). The mesoscale observing system outlined in this report will provide the backbone of data that will enable and accelerate the field of chemical weather prediction.
One important research activity will be the use of chemical data assimilation systems to help design the observing systems needed to produce better forecasts. We need to rigorously quantify the value added to a forecast by adding observations of additional species and above the surface, extending surface coverage; and adding and enhancing the utility of observations from satellites for chemical weather applications.
Only a small percentage of satellite observations is assimilated into prediction models. Debate continues about how to treat measurements of upwelling radiation. Preliminary studies at the UK Met Office have shown that soundings derived from the Atmospheric Infrared Sounder and the Infrared Atmospheric Sounding Interferometer, when inserted into computer prediction models, have had a greater impact on numerical predictions than the direct assimilation of radiance from those sensors, contrary to prevailing opinion. The reason may be that the radiance data are thinned both spatially and spectrally, whereas the derived soundings use all of the spectral information. Only further experimentation will resolve this issue.
The use in prediction models of cloud and hydrometeor information from satellites, surface-based ceiling observations, aircraft observations of clouds and icing, radar reflectivity, and lightning data is still primitive. More sophisticated assimilation techniques are sorely needed.
Correct specification of the statistical structure of model forecast errors is required for optimal performance of three- or four-dimensional variational data assimilation, especially the spatial covariance of model errors. The direct approach to this problem relies on an extensive network of dense observations for the direct calculation of differences between forecast and observed values and their means, standard deviations, and spatial covariances. Another approach is to estimate situation-dependent model errors by means of ensemble forecasts.
Further study is needed on how uncertainty in the initial state (primarily due to the sparsity of observations with respect to the grid resolution of today’s operational prediction models) translates into uncertainty in the model forecast.