Developments in Australian Doppler radar data classification and assimilation

The assimilation of radar data such as radial wind observations in high resolution NWP models requires automated, near real-time quality control. There are typically two or three stages of quality control prior to assimilation. The final, pre-assimilation stage may include comparison between observations and model values, and thinning/superobservation creation. The first stages may be labelled as ‘on-site' or ‘off-site' processing, and may differ for every radar network. The Australian operational radar network uses on-site raw signal processing for clutter filtering, noise correction and Signal Quality Index thresholding. Only reflectivity, velocity and sometimes spectrum width are transmitted for off-site processing. These parameters are then used to prepare the radar data for a range of applications including Quantitative Precipitation Estimation and Forecasts, wind profiles, and assimilation.

Most weather radar applications require the separation of precipitation and non-precipitation echoes. However, for radial wind data assimilation, some types of clear air echoes may still be suitable. Australian radars see abundant insect echoes that appear to have no major velocity bias due to insect air speed (compared with observation error). These observations could substantially increase the availability of radial velocity data for assimilation. To assess the potential of these data, some different types of non-precipitation echoes must be distinguished. A mostly comprehensive list of the types of echo seen on Doppler radars includes precipitation (stratiform and convective), aerofauna (insects, birds, bats), chaff, smoke, ground and sea clutter from normal and anomalous propagation, and second-trip echo. However, it is at a minimum only necessary to distinguish precipitation and potentially useful clear air echo (insects and smoke) from the rest.

The objective of this paper is to show how the available information can be used to provide data quality control suitable for radial velocity data assimilation. A classification system is being developed using a Naïve Bayes Classifier, combined with some logic and thresholding. The classifier uses several texture fields derived from the available radar parameters. Radar data from across the network was manually classified to create a training dataset. The classifier works in two complimentary ways. Firstly potential classes are determined for each value based on logic rules. Secondly the classifier is used to distinguish which class is represented. Using classification flags, the precipitation and clear air can be handled separately in the assimilation system.