[R-sig-Geo] gstat error: singular matrix in function LDLfactor()

[Edzer J. Pebesma]

karl.sommer at dpi.vic.gov.au wrote:

>When performing ordinary kriging under gstat in R I received the following
>error after the program had stopped.
>
>---R output
>vert.ok <- krige(vert~1, em38a, SPDF, model=vert.fit)
>[using ordinary kriging]
>
>"chfactor.c", line 130: singular matrix in function LDLfactor()
>
>gstat caught an error that occurred in the matrix library,
>the reason for it was: singular matrix
>
>HINT: Read the manual at http://www.gstat.org/ ;
>look for: Trouble shooting -> Error messages -> From meschach
>
>"chfactor.c", line 130: singular matrix in function LDLfactor()
>Error in predict.gstat(g, newdata = newdata, block = block, nsim = nsim,  :
>      matrix library error: gstat: matrix library error: singular matrix
>---end R output
>
>after consulting the gstat manual I found the following reference:
>
>"These two error messages may occur (a) during simulation, when an
>observation falls almost exactly at a simulation location (b) when two
>observations occur at identical location occur and noaverage was defined in
>its data definition. Solution to (a): increase the value of zero, to (b):
>remove the noaverage. Read also the next section."
>
>My input file was converted from lat/long to UTM using spTransform.  The
>sampling density is quite high and it is possible that after conversion
>from latlong to UTM because of rounding some measurements fall on identical
>coordinates.  Is this a possible explanation for the error?
>  
>
Yes. Also near-identical locations in combination with variograms 
without a nugget effect may caus it.

>The GSTAT online manual suggests to remove "noaverage" or increase "zero"
>to overcome the problem. However, I have not figured out how to do this
>when running gstat under R.
>  
>
Either package gstat, but more likely package sp has a function zerodist 
which will find point pairs with (nearly) zero distances. You may use 
the indexes it returns to do something with these points (e.g. remove 
one of them).

>Should I use only a subset of the orignial data or are there other
>alternatives to overcome the problem?
>  
>
Yes. An alternative would be to average observations, if they are scalar 
data.
--
Edzer

>Regards
>
>Karl
>
>_________________________________
>Karl J Sommer,
>Department of Primary Industries,
>Catchment & Agriculture Services,
>PO Box 905
>Mildura, VIC, Australia 3502
>
>Tel: +61 (0)3 5051 4390
>Fax +61 (0)3 5051 4534
>
>Email:     karl.sommer at dpi.vic.gov.au
>
>_______________________________________________
>R-sig-Geo mailing list
>R-sig-Geo at stat.math.ethz.ch
>https://stat.ethz.ch/mailman/listinfo/r-sig-geo
>  
>