The contamination of 'data impact' in global models by rapidly growing mesoscale instabilities

This paper illustrates a caveat in the 'data impact' method, in which the influence of assimilating a specific set of observations on a numerical weather forecast is evaluated. The 'signal' of data impact is defined as the difference between two forecasts, which are identical except that one forecast withholds the data in question from the assimilation. While it is anticipated that a coherent signal from observations in the midiatitude storm track is propagated dynamically by a forecast model from the vicinity of the observation locations, the reality is that the signal may become contaminated by initially small instabilities in dynamically unrelated locations. The initial signal may be non-zero in these remote locations due to small differences between the two analyses arising from truncation errors in the data assimilation scheme and/or the model's truncated spectral basis. The notion that the dynamical signal is contaminated is corroborated by assessing the data impact in the Northern Hemisphere of one rawinsonde released over Antarctica. After just a few hours, an amplifying signal manifests itself in convective areas in the Tropics, and even in locations along the midlatitude storm track where moist instabilities exist. Rapid growth and upscale evolution from the mesoscale to synoptic scales is evident. We find that the evaluation of the efficacy of a given set of observations on weather forecasts of more than two days may be compromised by initially small instabilities, particularly for spectral models and data assimilation schemes. The effective time is expected to be shorter in the Tropics. Copyright (C) 2007 Royal Meteorological Society.