Looking for collaborators: I would like this R package to become a collaborative effort, and I am looking for collaborators and/or testers (everyone is welcome!). The idea would be to couple the package development it with at least two collaborative papers: a software note, and an application case highlighting the importance of variable selection and modelling paradigm when building species distribution models. If you are interested, please send me an email to blasbenito at gmail.com or a message on twitter to @blasbenito

sdmflow

Work in progress!!

This R package intends to facilitate the design and execution of scientific workflows oriented to model species distributions over space and past-time. It particularly focuses on:

• Providing a consistent modeling grammar.
• Reducing environment cluttering by providing ad-hoc object classes to store results.
• Fitting models according to the paradigm "use versus availability".
• Providing a didactic documentation in the form of vignettes and function help files.

Modeling grammar

It particularly intends to facilitate the design of SDM workflows by providing a consistent modeling grammar, as easy to remember as possible, in order to reduce the cognitive load produced by packages with large numbers of functions. This grammar is based on the idea that an SDM workflow is composed by a limited set of conceptual steps:

Note: most of the functions mentioned below are still a work in progress.

• 1. Variable preparation, performed by functions named v_{action}. For example, v_match would match the coordinate reference system, extension, resolution, and mask, of a set of rasters from different sources stored in the same folder, or v_read, which would read those rasters into a stack or a brick.
• 2. Ocurrence preparation, done with functions named o_{action}, such as o_reproject, which would reproject the presence data to the same reference system of the raster data, or o_prepare, which would prepare the training data files required to fit SDMs from the presence data and the raster variables.
• 3. Selection of predictors, denoted by functions named s_{whatever}, such as s_lower_cor or s_lower_vif, which help to reduce multicollinearity in a training dataset, or s_auto, which selects variables automatically based on an estimate of their potential effect.
• 4. Modeling stage, performed with functions named m_{whatever}, such as m_fit to fit different models, m_performance to evaluate models, m_importance to assess variable importance, m_project to project SDMs into different times or places, or m_ensemble and m_stack to ensemble different models of the same species or stack together models of different species.
• 5. Reporting stage, intended to facilitate the storage and communication of modeling results, and peformed with functions named r_{whatever}, such as r_metadata to generate metadata documentation for each outcome of a modeling workflow, or r_report to generate an Rmarkdown template reporting the outcomes of a particular model.

Each stage will be as well represented by a single function that can perform at once the most important steps of each stage, so a complete modeling workflow could be written as follows, once the pertinent parameters are filled:

v_auto(...) %>%     #variable preparation
o_auto(...) %>%     #occurrence preparation
s_auto(...) %>%     #automatic variable selection
m_auto(...) %>%     #automatic modeling and evaluation
r_auto(...)         #report generation

Object classes

Another important target of "sdmflow" is to reduce the cognitive load produced by a cluttered environment. To solve this issue the package will rely on three main object classes (well, named lists with a particular structure):

• sdmflow.environment stores raster data in 4D (latitude, longitude, value, variable) or 5D (the latter plus time)

• sdmflow.training stores the training data frame along with other relevant objects resulting from variable screening (i.e. VIF results), the names of the selected variables, and the model formulas to fit models downstream.

• sdmflow.model stores the results of a model, including the model object itself, assessment results, variable importance, raster projection of the model, etc.

Modeling paradigm

The package sdmflow is based on the "use versus availability" modeling paradigm, which assumes that the presence records somehow reflect how a species uses the available habitat, which is represented by the background data. This modeling method gives higher habitat suitability to environmental values that are rare but disproportionately used by the species, following the idea that an accumulation of presence records over abundant environmental values can be the result of random processes, while the accumulation of presence records on rare environmental values is a clear signal of habitat selection.

The main advantage of this methods comes from its reliance on background data. Unlike absence or pseudo-absence data, background data does not have any interpretation problems, and simply represent an comprehensive sampling of the environmental conditions of the study area. This package also includes an option to work with "restricted background", generally taken from the area that is accessible to the species by dispersal.

Comprehensive documentation

This package was born after years of teaching species distribution models for GBIF.es, and I would like its documentation and vignettes to be as comprehensive as possible, so the package can become a source of tools and knowledge at the same time.