Quantifying the performanceof techniques for identifying clutter

Clutter is still a major limiting factor when retrieving rainfall rates with radar. MeteoSwiss have recently (Germann et al, 2006) modified their clutter elimination algorithm; as a result, the number of pixels deemed to be cluttered was reduced by a factor 3 and the discrepancy in daily radar/gauge comparisons reduced from 4 dB to 2 dB. The recent availability of digital receivers enables more complex parameters to be implemented operationally. In this work we compare the efficiency of four parameters: PR (power ratio, the ratio of the fluctuating power to the DC power); CPA (clutter phase alignment, equivalent to power ratio but for amplitude); CI (clutter index, or absolute power difference; the mean change in the value of Z using a specified lag between pulses), and fourthly, the well known TEX parameter (the texture of Z measured by the standard deviation of Z over neighbouring gates). A Bayes classifier is used to estimate the most likely target class (clutter or precipitation) for each gate, using the various parameters defined above, based on likelihood function established for clutter and precipitation from calibration data sets. The evaluation is then achieved from independent data sets. We have carried out these tests using data obtained from the C-band operational network in the UK, which has a beamwidth of one degree and transmits either 22 or 44 pulses per each degree of azimuth.

The performance of PR and CPA are similar and markedly better than CI. The TEX parameter performs less well, but is somewhat improved if it restricted to three gates along the particular ray (RTEX), The best performance is found for PR3 (which is PR smoothed over neighbouring pixels by averaging over a 3 x 3 window) combined with RTEX. In this case the probability of detection (POD) of clutter is > 99.3% and the false alarm rate (FAR) is < 0.3%. Based on this remarkable performance, the PR3/TEX parameter is now being rolled out for the radars in the UK operational network where it is replacing the CI parameter used previously. The performance of the PR3/TEX technique is comparable to that reported for dual polarization techniques. The next step will be to test a combined approach involving both PR3 and dual polarization to evaluate its performance