Hi again,
I'm getting more confused regarding the "accepted" forms of detrending
data prior to kriging. I've used a GAM (package mgcv) to detrend my
target variable. The residuals from this 9th order polynomial are well
behaved (normal distribution, only mild heteroskedasticity). I realize
that unlike the nlme package, the GAM from mgcv does not account for the
locations of the data, so the predicted data may not be statistically
optimal, but it was unclear whether the nlme package could also fit such
a trend to the data ( i suspected that it could, I'm obviously not
entering the code correctly). Oddly enough, adding the trend back to the
kriged residuals produced a similar map that using universal kriging
did...I suspect that this is because the majority of the prediction area
involves a portion of the data trend which could probably be modelled
reasonably well as a linear trend....
I guess, I'm not sure if there is a "standard" as to measure
against...As I also struggle with the concept of stationarity at times,
I find it is easy to get quickly confused. Almost all of the variograms
I produce from these data sets (either the raw data, or the residuals in
the presence of a weak trend) are bounded (i.e reach a sill), although a
few behave oddly at very large distances (well past the range of the
variogram)...I've interpreted this as simply a major reduction in the
numbers of point pairs that are available to compute the semivariance,
but my overall impression is that the data could be considered as second
order or intrinsically stationary...
If anyone has any thoughts or advice, I'd appreciate hearing your opinions.
Thanks,
Dave
_______________________________________________
R-sig-Geo mailing list
[email protected]
https://stat.ethz.ch/mailman/listinfo/r-sig-geo
