Spatial adjustment of environmental predictors and raster stacks

This function allows the transformation of provided environmental predictors (in SpatRaster format). A common use case is for instance the standardization (or scaling) of all predictors prior to model fitting. This function works both with SpatRaster as well as with stars objects.

Usage

predictor_transform(
  env,
  option,
  windsor_props = c(0.05, 0.95),
  pca.var = 0.8,
  method = NULL,
  ...
)

Arguments

env: A SpatRaster object.
option: A vector stating whether predictors should be preprocessed in any way (Options: 'none', 'scale', 'norm', 'windsor', 'windsor_thresh', 'percentile' 'pca', 'revjack'). See Details.
windsor_props: A numeric vector specifying the proportions to be clipped for windsorization (Default: c(.05,.95)).
pca.var: A numeric value between >0 and 1 stating the minimum amount of variance to be covered (Default: 0.8).
method: As 'option' for more intuitive method setting. Can be left empty (in this case option has to be set).
...: other options (Non specified).

Value

Returns a adjusted SpatRaster object of identical resolution.

Details

Available options are:

'none' The original layer(s) are returned.
'scale' This run the scale() function with default settings (1 Standard deviation) across all predictors. A sensible default to for most model fitting.
'norm' This normalizes all predictors to a range from 0-1.
'windsor' This applies a 'windsorization' to an existing raster layer by setting the lowest, respectively largest values to the value at a certain percentage level (e.g. 95%). Those can be set via the parameter "windsor_props".
'windsor_thresh' Same as option 'windsor', however in this case values are clamped to a thresholds rather than certain percentages calculated on the data.
'percentile' This converts and bins all values into percentiles, e.g. the top 10% or lowest 10% of values and so on.
'pca' This option runs a principal component decomposition of all predictors (via prcomp()). It returns new predictors resembling all components in order of the most important ones. Can be useful to reduce collinearity, however note that this changes all predictor names to 'PCX', where X is the number of the component. The parameter 'pca.var' can be modified to specify the minimum variance to be covered by the axes.
'revjack' Removes outliers from the supplied stack via a reverse jackknife procedure. Identified outliers are by default set to NA.

Examples