Hi Jonathan,
> 1. I want to perform regression kriging (i.e. krig the residuals of a
> regression model between two variables thereafter adding the kriged
> residuals back to the predicted map). Can this be done in 'universal
> kriging' mode in gsat, using the predicted surface from the regression
> model as the drift? I would assume if this was the case, the mask would be
> the predicted variable (Y) at all locations and X=4 in the command line,
> for example would, be a column containing the predicted (Y) variable at the
> data locations.
No. The X=.. columns and the values in the corresponding mask maps
should
be those of the covariates, or regressors, not the predicted surface.
E.g.
in a simple regression,
Y = a + b X + e
the X values for data locations should go in the (single) X=... column,
and those for the prediction locations in the (single) mask map. Note
that
gstat does the regression and residual kriging in one step, called
universal kriging.
> 2. Is there any way I can look into the details of the cross-validation
> statistics? i.e. generate a bivariate plot of the errors, compute other
> evaluation statistics, etc.?
If I recall correctly (IIRC), observed and predicted values are written
along with spatial coordinates to a file defined by the command:
set output = 'xvalid.out';
or on the command line:
gstat -o xvalid.out whatever.cmd
You can analyse them with your favourite statistics (or GIS) package.
I don't think that the regressors are written out, but now I come to
think about it, they should.
--
Edzer
