[R-sig-Geo] [Gstat-info] LOCAL universal kriging with GLOBAL reg. coeff. estimation

Edzer Pebesma edzer.pebesma at uni-muenster.de
Thu Aug 21 13:07:29 CEST 2008


Good question, I'll include r-sig-geo as well.

actually the prediction equations you end up with are kind of funny, and 
I've never seen them written out. Two different covariance matrices 
being inversed.

Package gstat can do the two-step approach: global BLUE, then simple 
kriging of residual, but that ignores the correlation of both terms, 
when added.

I'm quite sure that if you do Gaussian simulation however with gstat, 
the trend is fitted globally (and simulated from the corresponding GLS 
mean & covariance), whereas the kriging is done locally, so that should 
result in realisations that have the variance you want. You just have to 
compute the average and variance, to get estimates of the kriging mean 
and variance you're looking for. Also, it might be computationally 
demanding, both in time and space (memory).
--
Edzer

Vasat, Radim wrote:
> Dear all,
>
> My question is how to performe LOCAL universal kriging with 
> estimation of regression coefficients from the whole area 
> (GLOBAL estimation) at once.
>
> For large data (I have more than 8000 records) is computationaly too 
> demanding to handle with all the data for UK prediction. Hence,
> I prefer to do it LOCALLY (namx=100), but the reg. coeff. 
> would be estimated only with limited data in this case (not all
> data are included).
>
> To estimate the reg. coeff. first (with BLUE=TRUE) and then perform 
> the simple kriging with beta equals to the reg. coeff. and nmax=100
> might be the 
> solution. But the reg. coeff. estimation error is not included into the
> final kriging variance in this case.
>
> Anybody knows how to join these two requirements (LOCAL UK and GLOBAL
> estinmation of reg. coeff.) into one prediction???
>
> I will appreciate any coments!
> Radim
>
> _______________________________________________
> Gstat-info mailing list
> Gstat-info at geo.uu.nl
> http://mailman.geo.uu.nl/mailman/listinfo/gstat-info
>   

-- 
Edzer Pebesma
Institute for Geoinformatics (ifgi), University of Münster,
Weseler Straße 253, 48151 Münster, Germany.  Phone: +49 251
8333081, Fax: +49 251 8339763  http://ifgi.uni-muenster.de/




More information about the R-sig-Geo mailing list