These results only apply to satellite estimates over land. In so far as the rain processes and also many of the rain algorithms are different over water, we can't assume that the error measures we get over land apply equally well to ocean-based estimates.
The "frequency bias" is the ratio of the number of gridboxes raining in the satellite estimates to the number of gridboxes raining in the verifying analyses. A rain threshold of 1 mm/d or 20 mm/d was used to compute the frequency bias. The "multiplicative bias" is the ratio of mean estimated to mean analysed rain. The perfect value for both types of bias is 1.
The 95% confidence intervals were estimated using a bootstrap (resampling) method. Over the 29-month period different algorithms had different periods of missing data, so the exact set of days represented in these plots differs for the various algorithms. It would have been much better to have used a set of common days, but when I tried to do this the dataset shrunk too much. That is the reason for the large 95% confidence intervals on some of the IR estimates, for example. I personally tend to prefer the box and whisker plots to the pooled values since they give a better idea on the typical (and range of) performance you could expect on a daily basis.
Beth Ebert
(e.ebert@bom.gov.au)
Bureau of Meteorology Research Centre, Melbourne, Australia
7 June 2005