FURTHER COMPARISON OF SPATIAL METHODS FOR PREDICTING SOIL-PH

Spatial prediction methods have been compared using a carefully and specially designed survey of soil pH. Outliers seriously affected the performance of all prediction methods, and were removed for the comparison. Interpolators, Laplacian smoothing splines, and intrinsic random functions all behaved problematically, and universal kriging using parameter estimates obtained by the novel method called restricted maximum likelihood (REML) was consistently best. The data set contained an apparently obvious trend, but prediction by universal kriging was not improved by including this trend. The inclusion of close pairs stabilized the prediction methods, but there was no dramatic improvement with REML universal kriging, the best method for this data set.