This document provides an introduction to species distribution modeling with R. Species distribution modeling (SDM) is also known under other names including climate envelope-modeling, habitat modeling, and (environmental or ecological) niche-modeling. The aim of SDM is to estimate the similarity of the conditions at any site to the conditions at the locations of known occurrence (and perhaps of non-occurrence) of a phenomenon. A common application of this method is to predict species ranges with climate data as predictors.
In SDM, the following steps are usually taken: (1) locations of occurrence of a species (or other phenomenon) are compiled; (2) values of environmental predictor variables (such as climate) at these locations are extracted from spatial databases; (3) the environmental values are used to fit a model to estimate similarity to the sites of occurrence, or another measure such as abundance of the species; (4) The model is used to predict the variable of interest across an the region of interest (and perhaps for a future or past climate).
We assume that you are familiar with most of the concepts in SDM. If in doubt, you could consult, for example, the book by Janet Franklin (2009), the somewhat more theoretical book by Peterson et al. (2011), or the recent review article by Elith and Leathwick (2009). It is important to have a good understanding of the interplay of environmental (niche) and geographic (biotope) space – see Colwell and Rangel (2009) and Peterson et al. (2011) for a discussion. SDM is a widely used approach but there is much debate on when and how to best use this method. While we refer to some of these issues, in this document we do not provide an in-depth discussion of this scientific debate. Rather, our objective is to provide practical guidance to implemeting the basic steps of SDM. We leave it to you to use other sources to determine the appropriate methods for your research; and to use the ample opportunities provided by the R environment to improve existing approaches and to develop new ones.
We also assume that you are already somewhat familiar with the R language and environment. It would be particularly useful if you already had some experience with statistical model fitting (e.g. the
glm function) and with spatial data handling as implemented in the packages
sp. To familiarize yourself with model fitting see, for instance, the Documentation section on the CRAN webpage and any introduction to R txt. For the
raster package you can consult these pages.
When we present R code we will provide some explanation if we think it might be difficult or confusing. We will do more of this earlier on in this document, so if you are relatively inexperienced with R and would like to ease into it, read this text in the presented order.
SDM have been implemented in R in many different ways. Here we focus on the functions in the
dismo and the
raster packages (but we also refer to other packages). If you want to test, or build on, some of the examples presented here, make sure you have the latest versions of these packages, and their dependencies, installed. If you are using a recent version of R, you can do that with:
install.packages(c('raster', 'rgdal', 'dismo', 'rJava'))
This document consists of 4 main parts. Part I is concerned with data preparation. This is often the most time consuming part of a species distribution modeling project. You need to collect a sufficient number of occurrence records that document presence (and perhaps absence or abundance) of the species of interest. You also need to have accurate and relevant environmental data (predictor variables) at a sufficiently high spatial resolution. We first discuss some aspects of assembling and cleaning species records, followed by a discussion of aspects of choosing and using the predictor variables. A particularly important concern in species distribution modeling is that the species occurrence data adequately represent the actual distribution of the species studied. For instance, the species should be correctly identified, the coordinates of the location data need to be accurate enough to allow the general species/environment to be established, and the sample unbiased, or accompanied by information on known biases such that these can be taken into account. Part II introduces the main steps in SDM: fitting a model, making a prediction, and evaluating the result. Part III introduces different modeling methods in more detail (profile methods, regression methods, machine learning methods, and geographic methods). In Part IV we discuss a number of applications (e.g. predicting the effect of climate change), and a number of more advanced topics.
This is a work in progress. Suggestions are welcomed.
Robert J. Hijmans and Jane Elith