Function to create an ensemble of multiple fitted models

Ensemble models calculated on multiple models have often been shown to outcompete any single model in comparative assessments (Valavi et al. 2022).

This function creates an ensemble of multiple provided distribution models fitted with the ibis.iSDM-package. Each model has to have estimated predictions with a given method and optional uncertainty in form of the standard deviation or similar. Through the layer parameter it can be specified which part of the prediction should be averaged in an ensemble. This can be for instance the mean prediction and/or the standard deviation sd. See Details below for an overview of the different methods.

Also returns a coefficient of variation (cv) as output of the ensemble, but note this should not be interpreted as measure of model uncertainty as it cannot capture parameter uncertainty of individual models; rather it reflects variation among predictions which can be due to many factors including simply differences in model complexity.

Usage

ensemble(
  ...,
  method = "mean",
  weights = NULL,
  min.value = NULL,
  layer = "mean",
  normalize = FALSE,
  uncertainty = "cv",
  apply_threshold = TRUE
)

# S4 method for ANY
ensemble(
  ...,
  method = "mean",
  weights = NULL,
  min.value = NULL,
  layer = "mean",
  normalize = FALSE,
  uncertainty = "cv",
  apply_threshold = TRUE
)

Arguments

...: Provided DistributionModel or SpatRaster objects.
method: Approach on how the ensemble is to be created. See details for available options (Default: 'mean').
weights: (Optional) weights provided to the ensemble function if weighted means are to be constructed (Default: NULL).
min.value: A optional numeric stating a minimum value that needs to be surpassed in each layer before calculating and ensemble (Default: NULL).
layer: A character of the layer to be taken from each prediction (Default: 'mean'). If set to NULL ignore any of the layer names in ensembles of SpatRaster objects.
normalize: logical on whether the inputs of the ensemble should be normalized to a scale of 0-1 (Default: FALSE).
uncertainty: A character indicating how the uncertainty among models should be calculated. Available options include "none", the standard deviation ("sd"), the average of all PCA axes except the first "pca", the coefficient of variation ("cv", Default) or the range between the lowest and highest value ("range").
apply_threshold: A logical flag (Default: TRUE) specifying whether threshold values should also be created via "method". Only applies and works for DistributionModel and thresholds found.

Value

A SpatRaster object containing the ensemble of the provided predictions specified by method and a coefficient of variation across all models.

Details

Possible options for creating an ensemble includes:

'mean' - Calculates the mean of several predictions.
'median' - Calculates the median of several predictions.
'max' - The maximum value across predictions.
'min' - The minimum value across predictions.
'weighted.mean' - Calculates a weighted mean. Weights have to be supplied separately (e.g. TSS).
'min.sd' - Ensemble created by minimizing the uncertainty among predictions.
'threshold.frequency' - Returns an ensemble based on threshold frequency (simple count). Requires thresholds to be computed.
'pca' - Calculates a PCA between predictions of each algorithm and then extract the first axis (the one explaining the most variation).

In addition to the different ensemble methods, a minimal threshold (min.value) can be set that needs to be surpassed for averaging. By default this option is not used (Default: NULL).

Note by default only the band in the layer parameter is composited. If supported by the model other summary statistics from the posterior (e.g. 'sd') can be specified.

Note

If a list is supplied, then it is assumed that each entry in the list is a fitted DistributionModel object. Take care not to create an ensemble of models constructed with different link functions, e.g. logistic vs log. In this case the "normalize" parameter has to be set.

References

Valavi, R., Guillera‐Arroita, G., Lahoz‐Monfort, J. J., & Elith, J. (2022). Predictive performance of presence‐only species distribution models: a benchmark study with reproducible code. Ecological Monographs, 92(1), e01486.

Examples

# Method works for fitted models as well as as rasters
r1 <- terra::rast(nrows = 10, ncols = 10, res = 0.05, xmin = -1.5,
 xmax = 1.5, ymin = -1.5, ymax = 1.5, vals = rnorm(3600,mean = .5,sd = .1))
r2 <- terra::rast(nrows = 10, ncols = 10, res = 0.05, xmin = -1.5,
 xmax = 1.5, ymin = -1.5, ymax = 1.5, vals = rnorm(3600,mean = .5,sd = .5))
names(r1) <- names(r2) <- "mean"

# Assumes previously computed predictions
ex <- ensemble(r1, r2, method = "mean")

terra::plot(ex)