[R-sig-Geo] FW: Interpolcation option: IDW or OK?

[Tomislav Hengl]

Dear Yong Li,

I hope you will not mind me joining this interesting discussion. 

If there is no evident spatial auto-correlation structure (pure nugget effect), IDW/OK are as good
as randomly drawing a value from the global (normal) distribution. You can even test this using
cross-validation! In principle, there is no justification to use distance-based interpolators if
there is no evident spatial auto-correlation structure (maybe only the moving-window kriging, as
implemented in e.g. Vesper, or stratified kriging techniques could discover some local spatial
dependence). In addition, IDW should be considered an outdated technique, applicable only for
situations where the variogram is close to linear (e.g. elevation data and similar smooth surfaces).

What you should really consider using are the globaly available free maps/images (e.g. MODIS EVI,
SRTM DEM parameters etc.), and then see if you can explain some of the variability in your target
variable. 

But there will always be situations (especially in DSM applications) where you simply can not
explain much of the target variability, neither with auxiliary maps nor with spatial
auto-correlation. What to do then? I guess you simply have to collect more samples / more auxiliary
maps and then try again.

HTH

T. Hengl

See also:

Compendium of Global datasets:
http://spatial-analyst.net/wiki/index.php?title=Global_datasets

Regression-kriging:
http://spatial-analyst.net/wiki/index.php?title=Regression-kriging

Pebesma, E., 2006. The Role of External Variables and GIS Databases in Geostatistical Analysis.
Transactions in GIS, 10(4): 615-632.
http://dx.doi.org/10.1111/j.1467-9671.2006.01015.x 


> -----Original Message-----
> From: r-sig-geo-bounces at stat.math.ethz.ch [mailto:r-sig-geo-bounces at stat.math.ethz.ch] On Behalf
> Of Edzer Pebesma
> Sent: Monday, February 09, 2009 9:08 AM
> To: Yong Li
> Cc: r-sig-geo at stat.math.ethz.ch
> Subject: Re: [R-sig-Geo] FW: Interpolcation option: IDW or OK?
> 
> Yong Li wrote:
> > Hi Edzer,
> >
> > I would say the spatial structure is regarded not significant when c0/c0+c1 is very much greater
> than 75%. In my case I used even distance intervals and calculated c0/c0+c1 for log(OLSENP)
> greater than 85%. I knew this index sometimes is very fragile, very much depending on how we fit
> the model.
> >
> > However when I zoomed in by using variable distance intervals
> (boundaries=c(100,200,300,400,600,900,1000,1500,2000))and maxdist=2000 meters, I found a pretty
> good model-fitted experimental variogram. But the local OK interpolation using such a fitted model
> did not make sense when compared the predictions to the observations as in most areas values of
> OLSENP were severely underestimated. You may have seen my code with which I have tried the nested
> models, but unfortunately no luck either. I maybe think the parameters for local ordinary kriging
> are not optimized, but I have tried lots of sets of nmin, nmax and maxdist and did see the hopeful
> end.
> >
> > The journal editor insists in OK being better than IDW. I need to collect my evidence to defend
> my IDW choice. That is my intention raised such a question in our forum here.
> >
> I cannot find evidence in your data for such a claim; the cross
> validation statistics (rmse) seem to favour OK with your nested model.
> 
> In your first email, you stated the following:
> >> Normally if we do not find a significant spatial structure for a soil
> >> variable, we may choose IDW or other methods.
> What is the argumentation behind this? Who claimed this?
> 
> --
> Edzer Pebesma
> Institute for Geoinformatics (ifgi), University of Münster
> Weseler Straße 253, 48151 Münster, Germany. Phone: +49 251
> 8333081, Fax: +49 251 8339763 http://ifgi.uni-muenster.de/
> http://www.springer.com/978-0-387-78170-9 e.pebesma at wwu.de
> 
> _______________________________________________
> R-sig-Geo mailing list
> R-sig-Geo at stat.math.ethz.ch
> https://stat.ethz.ch/mailman/listinfo/r-sig-geo