Skip to contents

Add a control to a BiodiversityModel object to train ensembles of small models

Source: R/add_control_esm.R
add_control_esm.Rd

Estimating species distributions with few occurrences can be challenging. This is however often the case for many rare species, where seldom more than 5-10 occurrences are observed, or species that are estimated over a relatively coarse grid (e.g., aggregating observations over a coarse grid). In these cases, it is often that more complex algorithms tend to overpredict/overfit any given distributions (Breiner et al. 2015, 2018). A solution to this is to use ensembles of small models (ESM) that are essentially multiple separate models fitted on a subset of the covariates supplied. After inference, each model is then combined via ensemble() to produce a final distribution and coefficients.

In a range of studies it has been shown that this is an effective way to model species with few occurrences, and that it can outperform more complex models (Breiner et al. 2018, Erickson and Smith 2023).

See Details for the implementation in this package.

Usage

add_control_esm(x, n_covs = 2)

# S4 method for class 'BiodiversityDistribution'
add_control_esm(x, n_covs = 2)

Arguments

x

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

n_covs

A numeric on the number of covariates in each small model (Default: 2).

Value

Adds a control option to a distribution object that introduces the modelling of ensembles of small models in train().

Details

To make ensembles of small models (ESM) work, the distribution object gets assigned a flag that indicates that the model should be trained as an ensemble of small models within train(). The number of variables in each small model can be controlled via the n_covs parameter.

After inference and prediction, the coefficients from all models are extracted and also combined in an ensemble. A dummy engine_glm() model is then trained with the exact supplied coefficients per covariate. Note that this necessarily implies that the coefficients are not directly comparable to those of a single model. Further only linear modelling is supported to (re)capture the coefficients.

Note

Apart from ensembles of small models, consider also filtering the predictors prior to supplying them a distribution object. This can be done using via the function predictor_filter().

References

  • Breiner, F. T., Guisan, A., Bergamini, A., & Nobis, M. P. (2015). Overcoming limitations of modelling rare species by using ensembles of small models. Methods in Ecology and Evolution, 6(10), 1210-1218.

  • Breiner, F. T., Nobis, M. P., Bergamini, A., & Guisan, A. (2018). Optimizing ensembles of small models for predicting the distribution of species with few occurrences. Methods in Ecology and Evolution, 9(4), 802-808.

  • Erickson, K. D., & Smith, A. B. (2023). Modeling the rarest of the rare: a comparison between multi‐species distribution models, ensembles of small models, and single‐species models at extremely low sample sizes. Ecography, 2023(6), e06500.

See also

train()

Examples

if (FALSE) { # \dontrun{
 x <- distribution(background) |>
   add_biodiversity_poipa(species) |>
   add_predictors(covariates) |>
   add_control_esm(n_covs = 2)
} # }