tidysdm: Leveraging the flexibility of tidymodels for species distribution modelling in R

Abstract

In species distribution modelling (SDM), it is common practice to explore multiple machine learning (ML) algorithms and combine their results into ensembles. In R, many implementations of different ML algorithms are available but, as they were mostly developed independently, they often use inconsistent syntax and data structures. For this reason, repeating an analysis with multiple algorithms and combining their results can be challenging. Specialised SDM packages solve this problem by providing a simpler, unified interface by wrapping the original functions to tackle each specific requirement. However, creating and maintaining such interfaces is time‐consuming, and with this approach, the user cannot easily integrate other methods that may become available. Here, we present tidysdm, an R package that solves this problem by taking advantage of the tidymodels universe. tidymodels provide standardised grammar, data structures and modelling interfaces, and a well‐documented infrastructure to integrate new algorithms and metrics. The wide adoption of tidymodels means that most ML algorithms and metrics are already integrated, and the user can add additional ones. Moreover, because of the broad adoption of tidymodels, new statistical approaches tend to be implemented quickly, making them easily integrated into existing pipelines and analyses. tidysdm takes advantage of the tidymodels universe to provide a flexible and fully customisable pipeline to fit SDM. It includes SDM‐specific algorithms and metrics, and methods to facilitate the use of spatial data within tidymodels. Additionally, tidysdm is the first software that natively allows SDM to be performed using data from different periods, expanding the availability of SDM for scholars working in palaeontology, archaeology, palaeobiology, palaeoecology and other disciplines focussing on the past.