From the middle 1990s local high density atmospheric motion vectors (AMVs) began to be generated from geostationary satellite imagery and to be used operationally in the Australian Bureau of Meteorology. It was soon realised that their benefit to numerical weather prediction (NWP) was dependent on quality control, error characterisation and data selection. Here, by comparing the AMVs with collocated radiosonde data, we show that quality control (QC) considerably reduces vector error. In many cases, QC also reduces the correlated error and its associated length scale. Modern operational data assimilation schemes require estimation of observation errors, and they assume either that the errors are uncorrelated or require knowledge of their correlated error and length scale. A new construct, the expected error (EE) now employed at the Bureau, has been introduced to assist in quality control and error characterisation of AMVs. In this approach each vector is characterised by an expected error value, the correlated error and the related length scale, in addition to the set of currently available error indicators . The expected error allows direct comparison of the accuracy of vectors, based on different image types, or from different producers. At present, comparison of accuracy error is a more difficult process. Provision of the correlated error and length scale permits better use of the vectors in data assimilation, and provides a firm basis for data thinning.
Marshall et al. (Tue,) studied this question.