Fast kriging of large data sets with Gaussian Markov random fields

Linda Werner and Ola Hössjer


Centre for Mathematical Sciences
Mathematical Statistics
Lund Institute of Technology,
Lund University,
2004

ISSN 1403-9338
Abstract:
Spatial data sets are analysed in many scientific disciplines. Kriging, i.e. minimum mean squared error linear prediction, is probably the most widely used method of spatial prediction. For data sets with many observations, computation time and memory requirement could be an obstacle for kriging. In this article we approximate a Gaussian field with a Gaussian Markov random field on a lattice to speed up calculations and decrease memory requirements. By using a bilinear interpolation at non-lattice locations the method is well suited for non-lattice data.
Key words:
spatial interpolation, GMRF, non-lattice data