[STATSGRASS] questions regarding GRASS/gstat

[Edzer J. Pebesma]

Thomas Adams wrote:

> I have been attempting to use gstat within GRASS and have had success 
> up to a point. I am attempting to estimate precipitation over a large 
> river basin (~450000 sq km) from monthly (not mean monthly) point 
> values. I'm working with a Lambert Conic Conformal grid region with a 
> ~5 km grid spacing (the x & y spacings are slightly different). My 
> orginal dataset has nearly 500 data points, but some were too close 
> together (actually coincident points with different IDs), so I wrote a 
> program to eliminate points that were too close, <7.5 km. I have 
> increased this threshold several times, to 50 km and still get a gstat 
> error when I attempt the ordinary kriging, which says that I have a 
> "matrix library error: singular matrix".
>
> I must be doing something wrong since my once nearly 500 station 
> locations is now ~100! One other note, is that gstat reports that some 
> of my stations lie outside my mask region (the river basin boundary), 
> but I don't see that this should be a problem, is it? 

Once you end up with a singular covariance matrix, it is pretty hard
(and gstat does not attempt) to find out the reason why it occurs.
Possible reasons are (as you said): duplicate observations -- gstat
can catch them by taking their average, using something like:

data(rainfall): 'rainfall', ... , average=1;

other reasons may be the variogram model you use: the Gaussian
model often leads to singular models, and some authors (M. Stein)
argue not to use it at all; other models, such as linear-with-sill
should be avoided in 2D or 3D situations, but gstat does not check
this.

Let me know if this didn't help.
--
Edzer