nnKrige: Prediction at new location using neural network and modeling...
In martindurocher/floodRFA: A R-package for regional frequency of floods

Description Usage Arguments Value See Also Examples

The method uses neural network and bagging to estimate a trend and then correct for spatial correlation in the residual. The method uses a single hidden layer network that is carried out by nnet. Initially, the input variables are transformed using a SVD decomposition and the trend is obtained by aggregating (average) several bootstrap samples. The residuals are further predicted using kriging or thin plate spline.

1	nnKrige(form, x, xnew, n, k = 20, model = "none", loc, ...)

`form`	Formula defining the trend.
`x`	Data for training the model
`xnew`	Data at new locations.
`n`	Number of units
`k`	Number of bootstrap sample
`model`	Model for the residuals. Either 'tps' or variogram models (see vgm).
`loc`	Formula providing the spatial coordinates
`...`	Other argument pass to autoKrige

`pred`	Prediction at new locations.
`vgm`	Sample variogram.
`model`	Fitted variogram model. See vgm.
`resid.se`	Standard deviations of the kriging.

gamKrige, roiKrige.

## Gather info in on data.frame
 xd <- cbind( l1 = log(sapply(floodStream, mean)),
             floodVars, lon = floodCoord[,1], lat = floodCoord[,2])
 nsite <- nrow(xd)

 ## identify a validation and a training set
 valid <- seq(nsite) %in% sample(seq(nsite), round(.2*nsite))
 train <- !valid

 l1Form <- l1 ~ area + slope + elev + map + len + wb

 fit <- nnKrige(l1Form, x = xd[train,], xnew = xd[valid,], n = 30,
                model = c('Exp','Gau','Sph'), loc = ~lon+lat)
 print(fit)
 plot(fit)
 predict(fit)

 fit <- nnKrige(l1Form, x = xd[train,], xnew = xd[valid,], n = 30,
                model = 'tps', loc = ~lon+lat)

 fit <- nnKrige(l1Form, x = xd[train,], xnew = xd[valid,], n = 30)