This study investigates and quantifies the spatial error correlations of random errors in Atmospheric Motion Vectors (AMVs) derived by tracking structures in consecutive images from geostationary satellites. A good specification of the observation error is essential to assimilate any kind of observation in Numerical Weather Prediction. In most assimilation systems observation errors are assumed to be spatially uncorrelated. For AMVs, the height assignment or quality control procedures are likely to introduce spatially correlated errors.

The spatial structure of the error correlations is investigated based on a one-year dataset of pairs of collocations between AMVs and radiosondes. The error correlations for AMVs are obtained over dense sonde networks by assuming spatially uncorrelated sonde errors. Results for operational IR and WV wind datasets from METEOSAT-5 and 7, GOES-8 and 10, and GMS-5 are presented.

Winds from all five datasets show statistically significant spatial error correlations for distances up to about 800 km, with little difference between datasets, channels, or vertical levels. AMVs thus invalidate the assumption of spatially uncorrelated observation errors inherent in many data assimilation systems. The correlations tend to exhibit anisotropic structures with, for instance, longer correlation scales in North-South direction for the v-wind component. The AMV error correlations show some similarities with correlations of short-term forecast errors. The study reveals some shortcomings in the current use of AMVs in data assimilation, and the implications of the findings on the use of AMVs in data assimilation are discussed.