[GRASSLIST:9667] Re: cross validation on v.srf.rst. Does it work?

Markus Neteler neteler at itc.it
Thu Dec 29 06:48:30 EST 2005


Hi,

here response to an earlier question about the cross validation
feature in v.surf|vol.rst:

 http://thread.gmane.org/gmane.comp.gis.grass.user/9181

In 6.1-CVS the messages of both modules have been
syncronized to be less confusion (and easier to translate).

Based on the Slovakia3d dataset from
 http://mpa.itc.it/grasstutor/data_menu2nd.phtml
the following example has been added:

Example (Slovakia3d dataset):
 v.info -c precip3d
 v.vol.rst -c input=precip3d wcolumn=precip segmax=700 cvdev=cvdevmap tension=10
 v.db.select cvdevmap
 v.univar cvdevmap col=flt1 type=point

>From the results, parameters have to be optimized. It is recommended
to plot the CV error as curve while modifying the parameters.

The best approach is to start with tension, smooth and zmult with
rough steps, or to set zmult to a constant somewhere between
30-60. This helps to find minimal RMSE values while then finer steps
can be used in all parameters. The reasonable range is
tension=10...100, smooth=0.1...1.0, zmult=10...100.

In v.vol.rst the tension parameter is much more sensitive to changes
than in v.surf.rst. Usually tension=10...20 provide best results. But
the user should always check the result by visual inspection,
sometimes CV does not provide the best results, especially when the
density of data are insufficient. Then the optimal result found by CV
is an oversmoothed surface.

It is recommended to read
 Hofierka J., Parajka J., Mitasova H., Mitas L., 2002,
 Multivariate Interpolation of Precipitation Using Regularized Spline
 with Tension. Transactions in GIS  6, pp. 135-150. 

This hopefully helps.

Cheers

 Markus

-- 
Markus Neteler     <neteler itc it>       http://mpa.itc.it
ITC-irst -  Centro per la Ricerca Scientifica e Tecnologica
MPBA - Predictive Models for Biol. & Environ. Data Analysis
Via Sommarive, 18        -       38050 Povo (Trento), Italy




More information about the grass-user mailing list