virtualspecies, an R package to generate virtual species distributions

Abstract

virtualspecies is a freely available package for R designed to generate virtual species distributions, a procedure increasingly used in ecology to improve species distribution models. This package combines the existing methodological approaches with the objective of generating virtual species distributions with increased ecological realism. The package includes 1) generating the probability of occurrence of a virtual species from a spatial set of environmental conditions (i.e. environmental suitability), with two different approaches; 2) converting the environmental suitability into presence–absence with a probabilistic approach; 3) introducing dispersal limitations in the realised virtual species distributions and 4) sampling occurrences with different biases in the sampling procedure. The package was designed to be extremely flexible, to allow users to simulate their own defined species–environment relationships, as well as to provide a fine control over every simulation parameter. The package also includes a function to generate random virtual species distributions. We provide a simple example in this paper showing how increasing ecological realism of the virtual species impacts the predictive performance of species distribution models. We expect that this new package will be valuable to researchers willing to test techniques and protocols of species distribution models as well as various biogeographical hypotheses.

Concerns about the global decline of biodiversity in the face of global changes have led researchers to rely increasingly on species distribution models (SDMs) to predict current and future ranges of species. SDMs have thus become a central tool in conservation studies, to assess impacts of global changes on threatened species (Leroy et al. 2013), protected areas (Leroy et al. 2014), invasive species (Bellard et al. 2013), and also to test ecological hypotheses, such as niche conservatism (Strubbe et al. 2013). As a consequence, SDMs are constantly subject to methodological improvements regarding new techniques, protocols or evaluation methods (Elith et al. 2006, 2010, Phillips et al. 2009, Barbet‐Massin et al. 2012). These methodological improvements need testing and validation, which have mostly been based on empirical data. However, the use of empirical data during this validation phase is problematic because each dataset has many confounding factors, which preclude generalisation (Miller 2014). A valuable alternative is the simulation of virtual species distributions, because underlying mechanisms that generate such distribution patterns are known and can be manipulated independently (Hirzel et al. 2001, Meynard and Quinn 2007, Barbet‐Massin et al. 2012, Thibaud et al. 2014). The principle is to simulate the species’ probability of occurrence (hereafter called environmental suitability) with respect to one or several environmental gradients, and project it into a real or simulated landscape. This approach allows comparing the known ‘true’ distribution of the virtual species with the distribution predicted from the models, and to test independently the effects of each confounding factor (e.g. sampling scheme, model type, response shape). Consequently, the simulation of virtual species distributions is increasingly applied and advocated (reviewed by Miller 2014).

Recently, several software packages have been devoted to the simulation of virtual species. These packages can be used to test SDM outputs under multiple conditions. For instance, packages like RangeShifter (Bocedi et al. 2014) or HexSim (Schumaker 2015) are designed to simulate spatially explicit population dynamics. However, these software do not allow to simulate species–environment relationships, which is one of the most important issues in SDMs. As a result, they cannot be used to test the performance of SDMs to reproduce the simulated relationship or the influence of such relationships on model performance and classification rates, which are important limitations if we want to relate modelling to ecological properties of real species. The only software specifically developed for this purpose, SDMvspecies (Duan et al. 2015), is very limited in the number of possible species–environment relationships (i.e. linear, truncated linear or Gaussian response curves, Kong et al. 2014), and does not permit the users to customise occurrence–environment relationships (e.g. an ecophysiologist may want to use specific thermal performance functions to define a probability of occurrence for a virtual species, and this would not be feasible in SDMvspecies). In addition, this software does not include the possibility to simulate dispersal biases, or to sample species occurrences. As a result of the lack of suitable software, researchers modelling virtual species are left with the option to develop their own procedures to simulate species environmental relationship. In addition to being time consuming, it also means forgoing standardization (see, for example, the variety of methods described in Table 1 of Miller 2014). For these reasons, we developed virtualspecies, an open‐source package for the R environment (R Core Team) designed to provide a complete framework to generate virtual species distributions that allow to generate virtual species through various species–environment relationships, and to take into account distribution and sample biases.

Table 1. Description of the functions included in the virtualspecies package. Many additional options and customisations are available for the functions, all of which are comprehensively documented in the help files of the package

Core functions
generateSpFromFun	Generates a virtual species suitability map by defining its response functions to environmental variables.
generateSpFromPCA	Generates a virtual species suitability map by defining its response to a principal component analysis of environmental variables.
generateRandomSp	Randomly generates a virtual species distribution, including the conversion to presence–absence. The random aspects can be customised.
convertToPA	Converts the environmental suitability of a virtual species into presence–absence.
limitDistribution	Introduces a bias in the distribution of a virtual species by limiting its distribution to a chosen area.
sampleOccurrences	Samples occurrence points of the virtual species, with or without biases.
Utility functions
formatFunctions	Helps the user to format and illustrate the response functions as a correct input for generateSpFromFun.
plotResponse	Plots the relationship between the species and the environmental variables. Provides a plot appropriate for the method used to generate the virtual species.
removeCollinearity	Analyses and (if chosen) removes the collinearity among the environmental variables.
synchroniseNA	Ensures that cells containing NAs are the same among all the layers of a raster stack. Useful when building a stack of environmental variable coming from different sources, e.g. when combining climate with land use data.
Response functions currently embedded in the package
linearFun	A linear function of the form ax + b
logisticFun	A logistic function of the form 1/(1 + exp((x – β)/α))
quadraticFun	A quadratic function of the form ax² + bx + c
custnorm	A normal function parameterised by the mean, extreme values, and percentage of area under the curve between extremes.
betaFun	The asymmetrical beta function (Oksanen and Minchin 2002): k(x – p1)^α (p2 – x)^γ

Our objective with this framework is to integrate the main methodological advances published on the simulation of virtual species distributions, to provide a robust, comprehensive and user‐friendly package. Specifically, this package will allow researchers to simulate virtual species distributions with increased ecological realism. By increased ecological realism we mean simulations of species–environment relationships which are closer to the real species–environment relationships. This implies the possibility for the user to define any type of species–environment relationship, of increasing the complexity of the environment (i.e. improving cases with only a few predictors (e.g. Varela et al. 2014) to a variety of predictors), and to use of a probabilistic approach to convert environmental suitability to presence–absence (Meynard and Kaplan 2013). Increased ecological realism also implies the possibility of generating biases analogous to real biases, both in the realised distribution of the species (e.g. distribution limited by species dispersal abilities) and in the sampling of occurrences. Hence, the package was designed to be very flexible to integrate these possibilities and their full customisation.

The package is structured around four major steps (Fig. 1): 1) generating virtual species’ environmental suitability from a spatial set of environmental conditions, with two different approaches (Meynard and Quinn 2007, Barbet‐Massin et al. 2012); 2) converting the environmental suitability into presence–absence with a probabilistic approach (Meynard and Kaplan 2013); 3) introducing dispersal limitations in the realised virtual species distribution and 4) sampling occurrences with different biases in the sampling procedure. The package also includes various utility functions such as a function to visualise the species–environment relationships (Table 1). Hereafter, we detail the functioning of the package, step by step, and then we detail a working example of the package, showing how increasing ecological realism in the generation of virtual species impacts the predictive performance of SDMs. A comprehensive tutorial for the package is available online at < http://borisleroy.com/en/virtualspecies >.

Figure 1

Illustration of the package framework. The grey boxes on arrows indicate the names of the functions.

Package description

Requirements and input data

virtualspecies requires a standard installation of R and four extension packages, all of which can be installed from the Comprehensive R Archive Network: raster (Hijmans 2015), ade4 (Dray and Dufour 2007), dismo (Hijmans et al. 2014) and rworldmap (South 2011).

The package is designed to generate virtual species distributions from spatial environmental datasets (Fig. 1). These environmental datasets are gridded spatial data in the ‘raster’ format of the R package raster. Specifically, virtualspecies uses RasterStack objects, i.e. piles of rasters with the same spatial extent and resolution, where each layer corresponds to an environmental variable. For example, the global climate dataset WorldClim (< www.worldclim.com >) can easily be imported into R to a RasterStack format. Each layer of the input RasterStack corresponds to an environmental variable and must be named accordingly with a unique name. There is no limit to the number of layers of the input RasterStack, except for the capacities of the user's computer.

Step 1 – generation of the virtual species’ environmental suitability

The basis of generating a virtual species distribution consists in simulating the environmental suitability of a species, i.e. simulating its response to different environmental gradients, such as climatic variables. To simulate the environmental suitability, virtualspecies proposes two approaches (Fig. 1, step 1): 1) define a response function (e.g. linear, logistic, quadratic, gaussian) to each environmental variable, and combine these responses to define the environmental suitability (function generateSpFromFun); or 2) generate a principal component analysis (PCA) of all the environmental variables, define a response to each of the first two principal components (axes) and combine these responses to define the environmental suitability (function generateSpFromPCA).

1) Define responses to each environmental variable

This approach was introduced by Hirzel et al. (2001) which generated a virtual species by defining response functions to different predictor variables (e.g. a linear increasing response to forest cover, or a Gaussian response to an elevation gradient), and then combined the responses (using a weighted average) to obtain a habitat suitability value. It is the most frequently used approach to generate virtual species distributions (Elith et al. 2005, Meynard and Quinn 2007, Elith and Graham 2009, Bombi and D'Amen 2012, Meynard and Kaplan 2013). This approach is implemented in the function generateSpFromFun. Several functions are embedded with virtualspecies, based on functions already used in the literature (Table 1). virtualspecies is also extremely flexible because any other function existing in R can be used as a response function to environmental conditions (such as the normal distribution: dnorm); response functions can also be entirely created by the user (see the virtualspecies tutorial, section 2.3. at < http://borisleroy.com/en/virtualspecies >).

2) Define suitability from a PCA of environmental conditions

Defining response functions independently to each environmental variable (approach 1) can lead to virtual species with unrealistic environment conditions (e.g. a species with incompatible optima, such as an annual mean temperature of 35°C and a mean temperature of the warmest month of 5°C). While such cases can easily be avoided when only a few variables are considered, it becomes much more difficult when numerous variables are considered, especially if multiple species are simulated. An alternative approach for such cases consists in generating a PCA of environmental conditions, and then defining responses to two axes of this PCA (Barve et al. 2011, Barbet‐Massin et al. 2012). This approach ensures that the combination of environmental conditions is matched in the real world for the simulated species. It is implemented in the function generateSpFromPCA. This function is currently limited to the usage of normal response functions to the first two axes, but future versions will also implement any type of response function.

Step 2 – conversion of environmental suitability into presence–absence

The classical approach consists in defining a threshold to convert environmental suitability into a binary map of presence–absence (Hirzel et al. 2001, Bombi and D'Amen 2012). This approach, however, does not simulate the random processes acting on species occupancies and will always lead to threshold responses, despite the previous generation of non‐threshold environment–occurrence relationships (Meynard and Kaplan 2012, 2013). Such unrealistic virtual species can provide misleading results regarding the ability of modelling techniques to predict species distributions, particularly when non‐threshold responses are of interest. An appropriate alternative consists in applying a probabilistic approach in which a logistic function is used first to convert environmental suitability into a probability of occurrence (Fig. 1, step 2). A subsequent random draw using the probability of occurrence determines whether a particular cell is turned into a presence or an absence. Therefore, a binomial experiment is run in each cell, with the probability of occurrence as the parameter. A cell with a probability of occurrence of 0.2 will be assigned a presence in 2 out of 10 cases under this probabilistic approach (for more details and examples see the online tutorial, section 4.1. at < http://borisleroy.com/en/virtualspecies > and Meynard and Kaplan 2013). This probabilistic conversion to presence–absence implies that repetitions of the conversion process will differ, each providing a valid realisation of the true species distribution map. However notice that this approach also provides the flexibility to simulate threshold as well as non‐threshold responses (Meynard and Kaplan 2012, 2013). It is implemented in the function convertToPA.

The importance of species prevalence (i.e. the proportion of sites in which the species occurs), and particularly of the relationship between species prevalence and sample prevalence (i.e. the proportion of samples in which the species has been found) has been demonstrated on SDMs (Meynard and Kaplan 2012). Hence, the function automatically calculates the species prevalence for the user. Alternatively, the user can also specify the desired species prevalence to the function, which will automatically determine an appropriate conversion curve.

Step 3 (facultative) – introduce a distribution bias

One of the most disputed assumptions of SDMs is the assumption that species are at equilibrium with their environment (Guisan and Thuiller 2005), i.e. the assumption that they occupy their full range of suitable environmental conditions. This assumption is disputed because species’ realised distributions are often assumed to be restricted to a subset of their potential distributions, because of constraints to dispersal, competition, or stochastic extinction events for example. As a consequence, testing how well modelling techniques perform when the equilibrium assumption is violated is an important contribution of virtual species (Saupe et al. 2012). Virtual species generated with this package can be used to test such assumptions. The principle is to simulate a realised distribution for the virtual species which will be a subset of its potential distribution generated at step 2. This distribution bias can be achieved by different ways. Within the package, the function limitDistribution provides several convenient ways to limit the distribution of the species (Fig. 1 step 3). This functions restricts species’ presences (defined at step 2) to a spatial area defined by the user, and thus precludes any presence in cells outside this area. For example, in Fig. 1 step 3, the virtual species’ distribution was restricted to continental Africa only. To define the restricting area, different methods can be used in limitDistribution (i.e. using country, region or continent names, spatial polygons or extents).

Another possibility for users is to dynamically simulate the dispersion of their virtual species by combining outputs from virtualspecies (environmental suitability generated at step 1) with other modelling platforms, such as the MIGCLIM R package (Engler et al. 2012) or RangeShifter (Bocedi et al. 2014). In Supplementary material Appendix 2, we detail an example where we simulate the dynamic dispersion of a virtual species in Great Britain by combining virtualspecies with RangeShifter.

Step 4 – occurrence sampling

The last step consists in sampling observed occurrences for the virtual species with the function sampleOccurrences. This function can be used to sample different types of species occurrence (‘presence–absence’ or ‘presence only’), either randomly or with different biases similar to actual sampling biases (Fig. 1, step 4). For example, it is possible to assign a probability of detection to the virtual species, given the impact of imperfect detection on SDM performance (Lahoz‐Monfort et al. 2014). This probability of detection can be weighted by environmental suitability, to simulate smaller populations in less suitable areas. An error probability can be defined, to simulate misidentifications (i.e. erroneous presence in cells where the species is absent). A sampling intensity bias can also be applied, to simulate over‐ or under‐sampled areas (Phillips et al. 2009).

Utility functions

virtualspecies also includes various utility functions (Table 1), such as functions to visualise the relationship between the species and its environment (plotResponse), to randomly generate a virtual species (generateRandomSp), and remove the collinearity among environmental variables (removeCollinearity). These functions are summarised in Table 1 and have many customisable parameters described in their associated help files.

Example of application: evolution of SDM performance with increasing ecological realism of virtual species

The generation of virtual species distributions is often aimed at testing modelling techniques and protocols, with the ultimate objective of transferring the results to real‐world species. It is therefore important to attempt generating virtual species resembling real‐world species. Hence, it is crucial to establish the degree of model performance overestimation when using virtual species with poor ecological realism. Here, we provide an example of this overestimation, using a very simple case study.

We built this example using previous works showing three frequent characteristics of real‐world species that can be applied to virtual species. First, as explained in step 2, because of the random processes acting on species occupancies, real‐world species are more likely to have a gradual response to the environment rather than a threshold response (Meynard and Kaplan 2012, 2013). Second, sampling procedures are scarcely perfectly randomised, and often there are strong disparities in sampling intensities among geographical areas (see, e.g. the higher sampling intensity in Germany and the United Kingdom in spider distribution maps in Appendix S6 of Leroy et al. 2014), which impacts the performance of species distribution models (Phillips et al. 2009). Third, real‐world species may not be at the equilibrium with their environment, as this has been shown for many invasive species (Václavík and Meentemeyer 2012).

Simulations

We generated a virtual species distribution on the basis of six bioclimatic variables (www.worldclim.org/bioclim) : mean diurnal range (bio2), max temperature of warmest month (bio5), min temperature of coldest month (bio6), precipitation of wettest month (bio13), precipitation of driest month (bio14), precipitation seasonality (bio15). These variables were downloaded from WorldClim (www.worldclim.org) at a resolution of 0.17 decimal degrees, and the geographical area was restricted to the western Palearctic (longitude between –15 and 65; latitude between 30 and 75). On the basis of these environmental variables, we generated a virtual species’ environmental suitability with a PCA approach, using the generateSpFromPCA function. We manually defined the Gaussian response functions to axes 1 (mean = 4, standard deviation = 3) and 2 (mean = –3, standard deviation = 2). The resulting species–environment relationship was illustrated in Fig. 2 using the plotResponse function of the package.

Figure 2

Graphical representation of the simulation of a virtual species’ environmental suitability with a principal component analysis (PCA) of six climate variables (acronyms detailed in main text). In the top left corner the projection of each input climate variables on the PCA is shown. Each point of the PCA corresponds to a pixel of the input raster of climate variables. The Gaussian responses of the species to each axis are illustrated next to their respective axis. An ellipse is drawn around the area where the pixel suitability is highest. Points are coloured according to their climate suitability values for the virtual species: red points correspond to pixels with the highest suitability, and yellow to grey points correspond to pixels with the lowest suitability, as shown in the legend.

From this single virtual species’ environmental suitability, we generated four cases of increasing complexity (see the R script in Supplementary material Appendix 1).

Case 1: the environmental suitability was converted to presence–absence using a threshold of 0.7 (above the threshold presence is attributed, below absence is attributed). 150 presence points were randomly sampled.

Case 2: the environmental suitability was converted to presence–absence with a logistic curve of parameters β = 0.7 (inflexion point) and α = –0.1 (steepness of the slope). 150 presence points were randomly sampled.

Case 3: same as case 2, except that a sampling bias was introduced, where Germany and the United Kingdom were 50 times more sampled than elsewhere. This sampling bias emulates a situation where these two countries have naturalist societies who collected more data locally on the target species.

Case 4: same as case 2, except that the distribution of the species was subsequently limited to Great Britain and Ireland, using the function limitDistribution of the package. 150 presence points were randomly sampled.

Species distribution models

We built species distribution models for each case, using the biomod 2 modelling R package (Thuiller et al. 2009). We used the six raw climatic variables as predictors (i.e. not the axes of the PCA). We applied a classical presence‐only protocol, using 10 runs of 1000 randomly sampled pseudo‐absences to calibrate generalised linear models with the default settings of biomod. We then projected the predicted environmental suitability maps (averaged across the 10 pseudo‐absence runs) of each case (Fig. 3) and calculated two performance metrics by comparing the true to the predicted distributions for each case: the relative area under the receiver operating characteristic curve (ROC, Fielding and Bell 1997, ranging from 0.5 (no skill) to 1 (perfect score)) and the true skill statistic (TSS Allouche et al. 2006, ranging from 0 (no skill) to 1 (perfect score)). To compute the TSS metric, the predicted suitability maps were converted into presence–absence maps using a threshold maximising the TSS value.

Figure 3

Maps of the (A) true and (B–E) predicted environmental suitability of the virtual species according to the different simulations: (B) conversion of environmental suitability to presence–absence with a threshold of 0.7; (C) conversion of environmental suitability to presence–absence with a logistic curve of parameters β = 0.7 (inflexion point) and α = –0.1 (steepness of the slope); (D) same simulation as in (C), but a sampling bias was introduced where Germany and the United Kingdom were 50 times more sampled than elsewhere; (E) same simulation as in (C), but the dispersal of the species was limited to Great Britain and Ireland.

Results

As expected, there was a progressive degradation of the ability of SDMs to correctly predict the environmental suitability (Fig. 3 B–D) as real‐world aspects were included in the virtual species. Interestingly, the virtual distribution generated with a threshold conversion at step 2 was very well predicted above the threshold (i.e. above an environmental suitability of 0.7), but nothing could be predicted below the threshold (Fig. 3B). On the other hand, when a probabilistic conversion was used to generate the virtual species, SDMs were less performant to predict areas of high environmental suitability; but had the ability to detect the lower environmental suitability of the virtual species (Fig. 3C). When a sampling bias was introduced in Germany and the United Kingdom, the predicted suitability was even more biased. As expected, predicted suitability values showed higher values than true environmental suitability in areas with climate similar to these two countries, and lower predicted values than true environmental suitability elsewhere, such as in Spain and in eastern Europe (Fig. 3D). When the species was not at equilibrium with its environment, the predicted suitability was, unsurprisingly, dramatically underestimated (Fig. 3E).

Regarding the performance of SDMs, when the species has a threshold response where it is always absent below a suitability threshold and always present above that threshold (case 1), SDMs generate near‐perfect predictions according to both ROC and TSS (Fig. 4). However when the presence–absence corresponds to a probabilistic response (cases 2–4) there is a strong drop in predictive performance. The applied sampling bias (case 3) did further decrease the predictive performance, although to a lesser extent. Finally, SDMs applied on the species which was not at equilibrium with its environment (case 4) yielded the most spectacular drop in predictive performance, compared to the other cases.

Figure 4

Boxplots of performance metrics of species distribution models for the different virtual species simulations. Each boxplot is based on the 10 pseudo‐absence runs for the considered virtual species. ROC: area under the receiver operating characteristic curve; TSS: true skill statistic. The TSS was evaluated on maps of presence–absence predicted using a conversion threshold maximising the TSS value.

Implications

This simple example clearly illustrated how virtualspecies can be used to simulate distributions with increasing complexity and ask questions related to SDM performance. The case study show here confirmed the importance of using a probabilistic conversion into presence–absence (Meynard and Kaplan 2013). Results also extend the conclusions of Meynard and Kaplan (2012) to a situation where presence‐only data are used, which was not tested in the previous study. Our package also allows adding further complexity to the simulations by incorporating dispersal limitations and sampling bias, two additional biases that, expectedly, also strongly impacted the predictive performance of SDMs.

Discussion

virtualspecies is the first package providing a full working framework to generate virtual species distributions, including the simulation of species–environment relationships, conversion into presence–absence, introduction of distribution biases and sampling of occurrences. In this package we combined the best methodological advances from the literature into this single framework. virtualspecies should therefore help researchers in generating species with increased ecological realism, for example by designing complex species–environment relationships, avoiding the mistake of a threshold conversion into presence–absence (Meynard and Kaplan 2013), or introducing biases in the sampling of occurrence which are similar to real sampling biases (Phillips et al. 2009). Furthermore, virtualspecies can be coupled with other population dynamics platforms such as RangeShifter (Bocedi et al. 2014), which enables further complexity in the modelling of virtual species (Supplementary material Appendix 2).

In this paper we provided a very simple example showing how the performance of SDMs can be altered when applying different real‐world aspects on virtual species. The four cases of our example were generated with only a few lines of code (Supplementary material Appendix 1), which clearly highlights the simplicity of generating multiple different cases with the virtualspecies R package. The combination of its simplicity of use and possibilities of customisation will allow ecologists to easily generate multiple virtual species distributions with a fine control over the different simulation parameters. For example, the function generateSpFromFun can use any user‐defined response function. As a consequence, researchers willing to test a very particular response function, specific to model organisms, will be able to use virtualspecies, such as the different thermal performance curves described in Angilletta (2006). This will allow researchers to test new hypotheses, for example regarding the distribution of actual species or the robustness of modelling techniques to unusual species–environment relationships. In addition, the possibility of controlling every simulation parameter is valuable when generating virtual species distributions to test the robustness of modelling techniques and protocols to particular aspects or biases (Miller 2014).

Another major contribution of virtualspecies is an enhancement of transparency, replicability and comparability of studies involving virtual species. Indeed, the generated virtual species can be stored on the hard disk drive and provided in online supplementary materials of articles. Likewise, all the parameters used to generate virtual species are stored in the package outputs (Table 2), and can be provided in articles. These parameters can then be used as inputs to generate the same virtual species (see e.g. the R script in Supplementary material Appendix 1 to reproduce the example of this article), including the random aspects if the users use the R function set.seed before their simulations.

Table 2. Steps of the framework, associated functions of the virtualspecies R package, and output objects

Step	Functions	Output
1 – Generation of the virtual species environmental suitability	generateSpFromFun	A virtualspecies object containing three elements:
	generateSpFromPCA	• approach: the approach used to generate the species (response or PCA).
	generateRandomSp	• details: detailed parameters used to generate the environmental suitability. For example, if a response approach was chosen, then all the response functions and their parameters are stored here.
		• suitab.raster: the raster (gridded map) of environmental suitability of the virtual species.
2 – Conversion of the environmental suitability to presence–absence	convertToPA	The input virtualspecies object appended with two elements:
	generateRandomSp	• PA.conversion: the parameters used to convert environmental suitability into presence–absence.
		• pa.raster: the raster (gridded map) of true presence–absence of the virtual species.
3 – Introduce a distribution bias	limitDistribution	The input virtualspecies object appended with two elements:
		• geographical.limit: the geographical constraints used to limit the distribution of the virtual species.
		• occupied.area: the raster (gridded map) of cells accessible to the species given the geographical constraints.
4 – Sample occurrences	sampleOccurrences	A list containing three (unbiased sampling) to four (biased sampling elements):
		• sample.points: a table (data.frame) containing the coordinates of sampled points, the actual occurrence of the species in these sampled points (present or absent), and the observed occurrence (present or absent).
		• detection.probability: the chosen probability of detection of the virtual species.
		• error.probability: the chosen probability to assign presence in cells where the species is absent.
		• bias: if a bias was chosen, then the type of bias and the biased sampling area will be included.

Given the importance of SDMs in ecology, biogeography and biological conservation, and the fact that many methodological aspects of SDMs still need improvement, we expect that virtual species will have a major positive impact in these fields. We also expect virtual species to be useful in other fields. For example, virtual species can provide insights to test issues in biogeography, such as niche conservatism, but also to test hypotheses at micro‐scales, such as the impact of climate change on the distribution of insects on plant surfaces.

To cite virtualspecies or acknowledge its use, cite this Software note as follows, substituting the version of the application that you used for ‘version 0’:

Leroy, B., Meynard, C. N., Bellard, C. and Courchamp, F. 2015. virtualspecies, an R package to generate virtual species distributions. – Ecography 38: 000–000 (ver. 0).

Acknowledgements

BL, CB and FC were funded under the FFII ERA‐NET BiodivERsA project. We thank Wilfried Thuiller for fruitful discussions and wise advices and Régis Gallon for beta‐testing the package. This is VIMS contribution no. 3466.

Supplementary material (Appendix ECOG‐01388 at < www.ecography.org/readers/appendix >). Appendix 1–2.