Hi I have a set of two dimensional data( measured on a grid). Initial analysis strongly suggests that there is directional anisotropy, so I've been trying to deal with this.
One way I've found is to fit the variogram model by maximum likelihood (or REML) using geoR. geoR seems to account for anisotropy by deforming the locations, according to the anisotropy parameters (direction of maximum range, and the anisotropy ratio). Cross--validation suggests that this performs well, relative to fitting various models to isotropic sample variogram. The thing is, I want to use gstat for kriging. The reason for this is that I want to do block kriging. However, gstat seems to account for anisotropy by adjusting the variogram range according to the anisotropy parameters. So it seems that geoR deforms the coordinates, whereas gstat "deforms" the variogram.... The upshot of this is that, when I use gstat to do punctual kriging using the parameters estimated in geoR, I get different results to what I get when I use geoR for kriging. A possible alternative, is to use geoR to predict the deformed locations, and then use gstat to do the kriging WITHOUT specifying the anisotropy parameter. If I do this then gstat and geoR give the same results for kriging. This suggests that there is no difference in the algorithms used for kriging between the two packages. This approach seems to be OK for doing punctual kriging... However, I want to do block kriging (square blocks). The approach I've adopted for punctual kriging seems to fall over here. This is because I would specify the block size (in gstat) in the deformed coordinate space, whereas I actually want the blocks to be specified in the original geographic coordinate space. Am I making sense? What a mess!!! Can any one suggest an alternative, or is this all a bit silly? cheers Nick