R: Predict from an object of class '"gstatModel"'

predict.gstatModel-method {GSIF}

R Documentation

Predict from an object of class `"gstatModel"`

Description

Predicts from an object of class gstatModel-class using new prediction locations. The function combines predictions by regression (e.g. GLM) and interpolation of residuals (kriging) via the Regression-Kriging (RK) or Kriging with External Drift (KED, also known as Universal Kriging) framework.

Usage

 
## S4 method for signature 'gstatModel'
predict(object, 
     predictionLocations, nmin = 10, nmax = 30, debug.level = -1, 
     predict.method = c("RK", "KED"), nfold = 5, verbose = FALSE, 
     nsim = 0, mask.extra = TRUE, block, 
     zmin = -Inf, zmax = Inf, subsample = length(object@sp), 
     coarsening.factor = 1, vgmmodel = object@vgmModel,
     subset.observations = !is.na(object@sp@coords[,1]), betas = c(0,1), extend = .5, ...)
## S4 method for signature 'list'
predict(object, 
     predictionLocations, nmin = 10, nmax = 30, debug.level = -1, 
     predict.method = c("RK", "KED"), nfold = 5, verbose = FALSE, 
     nsim = 0, mask.extra = TRUE, block, 
     zmin = -Inf, zmax = Inf, subsample = length(object@sp), ...)

Arguments

`object`	object of type `"gstatModel"`
`predictionLocations`	object of type `"SpatialPixelsDataFrame"` prediction locations (must contain all covariates from the model)
`nmin`	integer; minimum number of nearest observations sent to `gstat::krige`
`nmax`	integer; maximum number of nearest observations sent to `gstat::krige`
`debug.level`	integer; default debug level mode sent to `gstat::krige`
`predict.method`	character; mathematical implementation of the `gstat::krige` interpolation method with covariates: Regression-Kriging (RK) or Kriging with External Drift (KED)
`nfold`	integer; n-fold cross validation sent to `gstat::krige.cv`
`verbose`	logical; specifies whether to supress the progress bar of the `gstat::krige.cv`
`nsim`	integer; triggers the geostatistical simulations
`mask.extra`	logical; specifies whether to mask out the extrapolation pixels (prediction variance exceeding the global variance)
`block`	numeric; support size (block support for objects of type `"SpatialPixelsDataFrame"` is chosen by default)
`zmin`	numeric; lower physical limit for the target variable
`zmax`	numeric; upper physical limit for the target variable
`subsample`	integer; sub-sample point observations to speed up the processing
`coarsening.factor`	integer; coarsening factor (1:5) to speed up the processing
`vgmmodel`	object of class `data.frame` corresponding to the `gstat::vgm` variogram
`subset.observations`	logical; vector specifying the subset of observations used for interpolation
`extend`	numeric; fraction of the range for which the spatial domain should be extended when searching for observations for kriging
`betas`	numeric; vector of the beta coefficients to be passed to the `gstat::krige`
`...`	other optional arguments that can be passed to `gstat::krige` and/or `predict.glm`

Details

Selecting predict.method = "KED" invokes simple kriging with external drift with betas set at 0 (intercept) and 1 (regression predictions used as the only covariate). This assumes that the regression model already results in an unbiased estimator of the trend model.
If not speficied otherwise, subset.observations by default selects only obserations within the spatial domain (bounding box) of the predictionLocations plus 50% of the one third of the extent of the area (extend). In the case of spatial duplicates in 2D or 3D, subset.observations will automatically remove all duplicates before running kriging. All points in 3D that stand exactly above each other will be removed by default.
Predictions can be speed up by using a larger coarsening.factor e.g. 2 to 5, in which case the ordinary kriging on residuals will run at a coarser resolution, and the output would be then downscaled to the original resolution using splines (via the warp method). In the case of predict.method = RK, the kriging variance is derived as a sum of the GLM variance and the OK variance, which is statistically sub-optimal.

Note

Predictions using predict.method = "KED" (the default gstat setting) can be time consuming for large data set and can result in instabilities (singular matrix problems) if the search radius is small and/or if all covariates contain exactly the same values. Predictions using predict.method = "RK" on the other hand can be speed up, but will typically underestimate the prediction variance (taken as a simple sum of the regression and ordinary kriging variances). Compare to the "KED" variance that includes also a cross-term (see Hengl et al. 2007 for more details).

Author(s)

Tomislav Hengl, Gerard B.M. Heuvelink and Bas Kempen

References

Hengl T., Heuvelink G.B.M., Rossiter D.G., 2007. About regression-kriging: from equations to case studies. Computers and Geosciences, 33(10): 1301-1315.

Predict from an object of class "gstatModel"