The technique of kriging is widely known to be limited by its assumption of stationarity, and performs poorly when the data involve localized effects such as discontinuities or nonlinear trends. A Bayesian partition model (BPM) is compared with results from ordinary kriging for various synthetic discontinuous 1-D functions, as well as for 1986 precipitation data from Switzerland. This latter dataset has been analysed during a comparison of spatial interpolation techniques, and has been interpreted as a stationary distribution and one thus suited to kriging. The results demonstrate that the BPM outperformed kriging in all of the datasets compared (when tested for prediction accuracy at a number of validation points), with improvements by a factor of up to 6 for the synthetic functions. © The Geological Society of London.

Original publication

DOI

10.1144/GSL.SP.2004.239.01.13

Type

Journal

Geological Society Special Publication

Publication Date

01/12/2004

Volume

239

Pages

195 - 209