Hi Vasya,
You have to take into account, when doing bootstrap, that you are
dealing with dependent data. Standard bootstrap will destroy this
dependence, so you should use alternatives such as some kind of block
bootstrap (see e.g. Politis and Romano, 1994. The stationary bootstrap.
JASA, 89, 1303-1313). Alternatively, you may use the estimated
covariance matrix to try to reproduce the spatial variability (see e.g.
Francisco-Fernandez et al., 2012. Nonparametric methods for spatial
regression. An application to seismic events, Environmetrics, 23,
85-93). In any case, if the number of data is large enough (as it
seems), my recommendation would be to use some kind of cross validation
instead of bootstrap.
Note also that the linear regression can be integrated in the
kriging procedure (using e.g. gstat or geoR), thereby obtaining an
estimate of the variance of the final prediction.
Best Regards,
Ruben Fernandez-Casal
El 10/12/2014 7:03, Vasya Pupkin escribió:
Dear ladies and gentlemen, I don’t really know it this is a suitable platform
for asking such questions, but stackoverflow could not help me, so I am trying
my luck here:).
What I am doing: I am trying to make a soil erodibility map. I have soil
erodibility measurements at points and I do a linear regression analysis to
predict soil erodibility from predictors (morphological and remotely sensed).
Then I apply the regression equation to the predictor rasters and get a raster
of soil erodibility. Then I calculate residuals and extra/inter-polate with
ordinary kriging (global) using my own formula discribing the variogramm. Then
I summ up the map from regression equation and the interpolated rsiduals to
create the final map, thus mimiking regression kriging as you have suggested.
The regression was done in R.
Now I want to validate the final map with bootstrap as it is implimented in R "boot" package. In the madual
it says you can wrap any complicated model in the "boot" sequence. So I wanted to write this whole
regression+residual_kriging model in the boostrap in R. After googling I fould this RSAGA package that provides a
direct access to SAGA's functionality. And there is this "rsaga.geoprocessor" module which works like CMD for
SAGA, and will not return the results back to R environment for further calculations but will save them to a file, like
SAGA's CMD normally does. And there is this "pick.from.points" module which has sort of deeper integration
with SAGA's modules (not all though) and is what I think I need but I can't figure out how I can use my own variogram
formula and how I can return the results back to R environment for further calculations. So in other words I need to do
a regression and then predict the values and extrapolate residuals for excluded points, so the RSAGA module should
return the values of extrapolated residuals back to “boot” calculation.
What I want to do is something like this (R script):
boot(mydata,
lm( resp~predict, ...)+
pick.from.points( lm(resp~predict)$residuals, method = "krige", model =
"a+b^c",...))
Do you have any idea how it could be done? If not, could you please advise on
any other ways or anybody who could know? The approach – regression + residuals
kriging is critical, but not the software, so if there are any other packages
for doing this, please advice.
Best wishes,
Maksim.
[[alternative HTML version deleted]]
_______________________________________________
R-sig-Geo mailing list
R-sig-Geo@r-project.org
https://stat.ethz.ch/mailman/listinfo/r-sig-geo
_______________________________________________
R-sig-Geo mailing list
R-sig-Geo@r-project.org
https://stat.ethz.ch/mailman/listinfo/r-sig-geo
