[R-sig-Geo] gstat: choice of inverse distance power in idw() function

[Jon Olav Skoien]

I agree with Paul that kriging might be a better option, but if you want 
to use IDW, the intamap package includes a function for optimizing the 
power based on cross-validation. In the example for the function 
interpolate, replace the penultimate line with:
 x = interpolate(meuse, meuse.grid, list(mean=TRUE), methodName = "idw")
to see how it works.
Tweaking the power to get a smooth surface is not necessarily the best 
choice, unless you know how smooth the surface should be.

Cheers,
Jon

Paul Hiemstra wrote:
> Hi Erik,
>
> IDW does not include a formal interpretation of the inverse distance 
> power. You could seperate your dataset into a validation and 
> interpolation set and try different idp's and see which one preforms 
> best.
>
> Another option would be to use kriging. Kriging fits the spatial 
> dependence vs distance to the data. In my view this makes kriging, as 
> long as the assumptions are honored, a preferable approach. The 
> automap package provides easy acces to kriging by providing some 
> wrapper code around gstat.
>
> cheers,
> Paul
>
> On 08/30/2010 06:31 PM, Mudrak, Erika [EEOBS] wrote:
>> I used the gstat package to interpolate measurements of eight 
>> environmental variables in a square 15.4 m x 15.4 m, and then I used 
>> model selection from another package to build models of dependence of 
>> plant population locations on those environmental variables.  I used 
>> the idw() function to interpolate the environmental variables.  The 
>> model selection procedure defined which of the eight variables helped 
>> to explain the patterns seen in my plant populations.
>>
>> Are there any guidelines for the choice of the inverse distance 
>> weighting power (idp)?   I had been using idp=2, because it was the 
>> default, but for some variables it made the surface look not very 
>> smooth.   I have tried my models on surfaces with other values of 
>> idp, and changing this parameter causes the model selection procedure 
>> to arrive at different models.
>>
>> Does anyone have any advice or guidelines about the choice of the ipd 
>> parameter, other than "tweaking" it until the surfaces look smooth?
>>
>> Thank you,
>>
>> Erika Mudrak
>>
>>     [[alternative HTML version deleted]]
>>
>> _______________________________________________
>> R-sig-Geo mailing list
>> R-sig-Geo at stat.math.ethz.ch
>> https://stat.ethz.ch/mailman/listinfo/r-sig-geo
>>    
>
>