Comparative real data tests of some objective analysis methods by withholding observations

Several objective analysis methods were tested using 20 years of surface pressure observalions over southeastern Australia. The methods were assessed by withholding observations from a few stations, and comparing the analysed and observed values at the locations of the withheld stations. Best results were obtained by statistical interpolation using appropriate forms of correlation function. However the Cressman successive correction method was only slightly less accurate than statistical interpolation with the hest functions, and was better than statistical interpolation with a Gaussian correlation function. There is some suggestion that when the covariance structure is not well known, a modest improvement in statistical interpolation accuracy may be achievable by changing the 'noise to signal' parameter from case to case, using generalised cross validation.