LatticeKrig: Multiresolution Kriging based on Markov random fields
Functions for the interpolation of large spatial datasets.
This package follows a "fixed rank Kriging" approach using a
large number of basis functions and provides spatial estimates
that are comparable to standard families of covariance
functions. Using a large number of basis functions allows for
estimates that can come close to interpolating the observations
(a spatial model with a small nugget variance.) The covariance
model for this method can approximate the Matern covariance
family but also allows for a multi-resolution model and
supports efficient computation of the profile likelihood for
estimating covariance parameters. This is accomplished through
compactly supported basis functions and a Markov random field
model for the basis coefficients. These features lead to sparse
matrices for the computations. One benefit of this approach is
the facilty to do unconditional and conditional simulation of
the field for large numbers of arbitrary points. There is also
the flexibility for estimating nonstationary covariances.
Included are generic methods for prediction, standard errors
for prediction, plotting of the estimated surface and
conditional and unconditional simulation.