Re: Problems with gstat gaussian variogram and cross validation when fix nugget at 0
Fixing the nugget at zero for the Gaussian varigoram can often cause numerical problems in kriging. If you have two observations that are close in space, their modelled semivariance will be so small that you get numerical issues with the covariance matrix. In some cases it will be singular, in your case it could be inverted, but the absolute values of the weights will be very high for some of the cross-validation locations. It seems you have two stations that are very close to each other, if you remove one of these, you can see how the results will be more similar (still not the same) to the model with a small nugget effect.
Jon Olav Skøien
Joint Research Centre – JRC.E.1
Disaster Risk Management Unit
Building 26b 1/144 | Via E.Fermi 2749, I-21027 Ispra (VA) Italy, TP 267
[hidden email] Tel: +39 0332 789205
Disclaimer: Views expressed in this email are those of the individual and do not necessarily represent official views of the European Commission.
From: R-sig-Geo [[hidden email]] on behalf of Stefano Menichetti [[hidden email]]
Sent: 12 October 2018 16:36
To: [hidden email] Subject: [R-sig-Geo] Problems with gstat gaussian variogram and cross validation when fix nugget at 0
Dear all, I have a question about fixing nugget at 0 in a Gaussian
variogram using *gstat *package
I have encountered a serious problem on a variogram of piezometric data
that give "strange" results if I want to fix nugget to 0.