[STATSGRASS] krigged dem lost detail

[Roger Bivand]

On Wed, 15 Mar 2006, Markus Neteler wrote:

> Carlos,
> 
> I didn't follow this thread in details (although very interesting).
> There is cross validation for optimizing tension and smoothing,
> did you make use of that (sorry, if already written...)?

I think that Hamish and Markus are right: both rst and kriging have 
parameters which may lead to this result if set in particular ways. Maybe 
cross validation is a way of comparing their leave-one-out prediction 
performance for (for example) a subscene of the STRM data you've been 
using. 

You'd have to try a bit to find out which parameters might be the critical
ones. I have compared gstat kriging using CV with Tps in the fields
package, I have a presentation with the gory details (it was actually
comparing with GeoStatistical Analyst on the Meuse Bank data, gstat with
the flood frequency factor in R came out best).

Because upgrading 90m data is likely to be something a lot of people want 
to do, it would be worth trying to find out what is going on here. Can you 
post a suitable subscene and its location metadata - say as a geotiff?

Roger



> 
> I received a script from Jaro to automate the parametrization.
> Let me know if you need it (maybe it should go into the Wiki,
> so that you/others could document the procedure)
> 
> Markus
> 
> On Wed, Mar 15, 2006 at 07:16:11PM +1300, Hamish wrote:
> > > I am comparing the results we get when re-interpolating a DEM to a
> > > higher resolution. I did it with krigging (in R, with gstat) and with
> > > RST. In the final result it looks like the krigged DEM has lost
> > > detail, while the RST DEM kept those.
> > > see it:
> > > http://www.igc.usp.br/pessoais/guano/tempo/krigging.jpg
> > > http://www.igc.usp.br/pessoais/guano/tempo/rst.jpg
> > > 
> > > I don't know why this happens, since they were interpolated wit the
> > > same resolution, based on the same dataset.
> > 
> > 
> > be careful you haven't added detail that isn't really there due to
> > spline artifacts. Perhaps resample known DEM at lower resoution and then
> > re-interpolate at original resolution to compare which does a better
> > job recreating the original DEM?
> > 
> > r.resamp.rst
> > 
> > 
> > 
> > Hamish
> > 
> > _______________________________________________
> > statsgrass mailing list
> > statsgrass at grass.itc.it
> > http://grass.itc.it/mailman/listinfo/statsgrass
> 
> 

-- 
Roger Bivand
Economic Geography Section, Department of Economics, Norwegian School of
Economics and Business Administration, Helleveien 30, N-5045 Bergen,
Norway. voice: +47 55 95 93 55; fax +47 55 95 95 43
e-mail: Roger.Bivand at nhh.no