Skip to contents

Cross-validate a soil interpolation method

Source: R/kriging.R
cv_soil.Rd

Estimates the prediction error of an interpolation method by k-fold cross-validation: the points are split in `nfold` groups; each group is in turn removed, the model is re-fitted (including the variogram) on the remaining points and used to predict the removed ones.

With `block_size`, folds are made of square spatial blocks instead of random points (*spatial cross-validation*). This is more demanding: it mimics the prediction of areas located far from any sample, where interpolators that simply copy neighbours are penalised.

Usage

cv_soil(
  points,
  value,
  method = c("kriging", "kernel"),
  covariates = NULL,
  model = "Exp",
  log10 = TRUE,
  nfold = 10,
  block_size = NULL,
  seed = 1,
  radius = 200,
  size_std = 3,
  kappa = 0.5,
  cutoff = NULL,
  width = NULL
)

Arguments

points

An `sf` object of POINT geometries in a projected CRS (metres).

value

Character. Name of the concentration column (e.g. `"cd"`).

method

Character. `"kriging"` ([krige_soil()]) or `"kernel"` ([kernel_smooth_soil()]).

covariates

Optional `SpatRaster` of drift covariates (must cover the points and the output grid, same CRS).

model

Character vector of candidate variogram models (see [fit_soil_variogram()]); the best fitting one is used. Ignored when `vgm` is given.

log10

Logical. Krige `log10(value)` (default `TRUE`).

nfold

Integer. Number of folds (default 10).

block_size

Numeric. Side (m) of the spatial blocks; `NULL` (default) for random folds.

seed

Integer. Random seed for the fold assignment.

radius, size_std

Kernel standard deviation (m) and window half-size (in number of `radius`), for `method = "kernel"` (see [kernel_smooth_soil()]). For speed, the cross-validation evaluates the Gaussian kernel exactly at the left-out points instead of on a raster.

kappa

Matérn smoothness, passed to [fit_soil_variogram()].

cutoff, width

Passed to [soil_variogram()].

Value

A `data.frame` of class `soil_cv` with the observed value (`obs`), the prediction (`pred`) and the `fold` of each point (on the log10 scale if `log10 = TRUE`). Use [summary()] on it to get RMSE, MAE, bias and R2.

See also

[krige_soil()], [kernel_smooth_soil()]

Examples

# \donttest{
data(soil_metaleurop)
cv <- cv_soil(soil_metaleurop, "cd", model = "Exp")
summary(cv)
#>        RMSE       MAE         bias        R2   n
#> 1 0.2506923 0.1622391 0.0001610769 0.6025171 587
# }