Add predictors from a fitted model to a Biodiversity distribution object

This function is a convenience wrapper to add the output from a previous fitted DistributionModel to another BiodiversityDistribution object. Obviously only works if a prediction was fitted in the model. Options to instead add thresholds, or to transform / derivate the model outputs are also supported.

Usage

add_predictors_model(
  x,
  model,
  transform = "scale",
  derivates = "none",
  threshold_only = FALSE,
  priors = NULL,
  ...
)

# S4 method for BiodiversityDistribution
add_predictors_model(
  x,
  model,
  transform = "scale",
  derivates = "none",
  threshold_only = FALSE,
  priors = NULL,
  ...
)

Arguments

x

distribution() (i.e. BiodiversityDistribution) object.

model

A DistributionModel object.

transform

A vector stating whether predictors should be preprocessed in any way (Options: 'none','pca', 'scale', 'norm')

derivates

A Boolean check whether derivate features should be considered (Options: 'none', 'thresh', 'hinge', 'quad') )

threshold_only

A logical flag indicating whether to add thresholded layers from the fitted model (if existing) instead (Default: FALSE).

priors

A PriorList object. Default is set to NULL which uses default prior assumptions.

...

Other parameters passed down

Details

A transformation takes the provided rasters and for instance rescales them or transforms them through a principal component analysis (prcomp). In contrast, derivates leave the original provided predictors alone, but instead create new ones, for instance by transforming their values through a quadratic or hinge transformation. Note that this effectively increases the number of predictors in the object, generally requiring stronger regularization by the used Engine. Both transformations and derivates can also be combined. Available options for transformation are:

• 'none' - Leaves the provided predictors in the original scale.

• 'pca' - Converts the predictors to principal components. Note that this results in a renaming of the variables to principal component axes!

• 'scale' - Transforms all predictors by applying scale on them.

• 'norm' - Normalizes all predictors by transforming them to a scale from 0 to 1.

• 'windsor' - Applies a windsorization to the target predictors. By default this effectively cuts the predictors to the 0.05 and 0.95, thus helping to remove extreme outliers.

Available options for creating derivates are:

• 'none' - No additional predictor derivates are created.

• 'quad' - Adds quadratic transformed predictors.

• 'interaction' - Add interacting predictors. Interactions need to be specified ("int_variables")!

• 'thresh' - Add threshold transformed predictors.

• 'hinge' - Add hinge transformed predictors.

• 'bin' - Add predictors binned by their percentiles.

Examples

if (FALSE) {
 # Fit first model
 fit <- distribution(background) |>
        add_predictors(covariates) |>
        add_biodiversity_poipa(species) |>
        engine_glmnet() |>
        train()

 # New model object
 obj <- distribution(background) |>
        add_predictors_model(fit)
 obj
}