Using species distribution modelling to determine opportunities for trophic rewilding under future scenarios of climate change

Abstract

Trophic rewilding, the (re)introduction of species to promote self-regulating biodiverse ecosystems, is a future-oriented approach to ecological restoration. In the twenty-first century and beyond, human-mediated climate change looms as a major threat to global biodiversity and ecosystem function. A critical aspect in planning trophic rewilding projects is the selection of suitable sites that match the needs of the focal species under both current and future climates. Species distribution models (SDMs) are currently the main tools to derive spatially explicit predictions of environmental suitability for species, but the extent of their adoption for trophic rewilding projects has been limited. Here, we provide an overview of applications of SDMs to trophic rewilding projects, outline methodological choices and issues, and provide a synthesis and outlook. We then predict the potential distribution of 17 large-bodied taxa proposed as trophic rewilding candidates and which represent different continents and habitats. We identified widespread climatic suitability for these species in the discussed (re)introduction regions under current climates. Climatic conditions generally remain suitable in the future, although some species will experience reduced suitability in parts of these regions. We conclude that climate change is not a major barrier to trophic rewilding as currently discussed in the literature.

This article is part of the theme issue ‘Trophic rewilding: consequences for ecosystems under global change’.

1. Introduction

Trophic rewilding, the proposal to (re)introduce species to promote self-regulating biodiverse ecosystems, is a promising ecological restoration strategy [1]. Inspired by increasing evidence that large carnivores and herbivores are often important for ecosystem functioning and biodiversity maintenance [2–4], trophic rewilding aims to restore top-down trophic interactions and associated cascades [1]. Because trophic rewilding is a future-oriented proposal that seeks to learn from the past rather than recreate it, a key priority is to identify how threats in the Anthropocene, such as human-mediated climate change, may impact these ecological restoration projects [1,4].

Climate change threatens global biodiversity, ecosystem function and human well-being [5]. The Earth has warmed by about 0.74°C in the last 100 years and global mean temperatures are projected to increase further by 4.3 ± 0.7°C by 2100 [6]. Species have already responded to recent climatic shifts [7], and species redistribution is projected to accelerate into the future [5], increasingly driven by climate change [8]. Climate change will impact all biodiversity and conservation management [5], including for trophic rewilding projects [1].

A critical aspect in planning trophic rewilding projects is selection of suitable sites that match the biotic and abiotic needs of the focal species under current and future climates [1]. Species distribution models (SDMs; also known as ecological niche models, bioclimatic envelope models and habitat suitability models, among other names) are currently the main tools used to derive spatially explicit predictions of environmental suitability for species [9–11]. They typically achieve this by statistically relating environmental descriptors directly to species occurrence or abundance, although other approaches, including mechanistic or process-based models which explicitly link environmental conditions to species' physiological responses, also exist ([12,13]; hereafter mechanistic SDMs). SDMs have the potential to play a critical role in identifying suitable habitat for trophic rewilding under future scenarios of global change, including for population restoration, the restoration of declining or extirpated populations, and conservation introduction, the movement and release of an organism outside its indigenous range [1]. However, the extent of their adoption in aid of trophic rewilding projects has been limited.

Our aim is to provide a brief overview of the applications of SDMs to identify suitable sites for focal species of trophic rewilding projects under current and future climates. We do not propose an extensive review of SDMs, or undertake an exhaustive quantitative assessment of the grey literature, but based on focal species we emphasize the broad utility of SDMs to identify suitable habitat for trophic rewilding. For the selected focal species, we use SDMs to assess whether proposed regions are climatically suitable under current and future climates. Finally, we discuss methodological considerations needed when building and using SDMs to determine opportunities for trophic rewilding under global change.

2. Approaches to model suitable habitat to determine trophic rewilding opportunities

Different approaches can be used to project suitable habitat for species under global change scenarios. These approaches are typically placed in two classes: correlative and mechanistic SDMs (for more extensive discussion of ecological theory and methodology for correlative SDMs, see [9–11]; and for mechanistic SDMs, see [12–14]).

Distributions of species are typically estimated through the use of correlative SDMs that aim to represent the realized niche of a species [9,15]. Correlative SDMs can be built from presence-only data, presence/absence data or abundance data, either from fieldwork or specimens records, and often with the use of pseudo-presence locations, with spatial bias known to occur in all these kinds of distribution records [9,10,16,17]. The use of correlative SDMs to support conservation decision making has been encouraged, including to identify climate refugia, to identify potential reserve locations and to identify suitable habitat for the translocation of threatened species [18]. Advantages of correlative SDMs are that they are spatially explicit, applicable to a wide range of taxa at various spatial scales, and relatively quick and cheap to apply [10]. Like any model, correlative SDMs rely on assumptions and incorporate uncertainties (see for detailed discussions [19] and [20]). Assumptions particularly relevant for modelling the potential distribution of trophic rewilding candidates under climate change are that the full realized niche is captured, that the niche is not truncated due to anthropogenic influences (which is especially problematic for megafauna due to human impacts on their ranges), and that the same environmental conditions in time or space prevail in both areas or periods [19,20]. Primary sources of uncertainty include climate projections (which also apply to mechanistic SDMs, and arise from the differences in how general circulation models are built), algorithmic differences in methods and models as well as the selection of model predictors, and the extent to which the assumptions made about a species' biology are appropriate [19,20].

Recent advances to predict future habitat suitability include mechanistic SDMs that explicitly integrate physiology, demographics, dispersal and biotic interactions [12,13]. Mechanistic SDMs simulate processes hypothesized via rules or equations that have biologically meaningful parameters to represent the fundamental niche of a species [13]. While mechanistic SDMs are considered by some authors to be more robust and theoretically defensible than correlative SDMs for predicting species' responses to global change, extensive application is currently precluded by the fact that they require detailed data that are lacking for most species [14]. The main sources of uncertainty in mechanistic SDMs relate to model parameters, and to combining data collected at different spatial resolution [12,13]. Mechanistic SDMs, for example, can overcome traditional correlative SDM limitations for species that have naturally restricted distributions [21] and when making predictions under novel climates [22].

3. Case studies

We present case studies of trophic rewilding candidates that have already been discussed in the published literature and which represent different continents and habitats (see table 1 for the list of species, their functional traits, their known or potential ecological roles, and the types of translocations they have been proposed for, and the electronic supplementary material for more detailed descriptions). Given the importance of large animals in trophic cascades and their widespread losses our focus is on megafauna, which are often defined as animals with adults larger than some threshold mass. Here, we use the trophic-function-based megafauna grouping of Malhi et al. [4] to select focal species, where megafauna on continents are defined as large herbivores (45–999 kg), megaherbivores (≥1000 kg), large carnivores (21.5–99 kg) and megacarnivores (≥100 kg). Large herbivores and megaherbivores shape the structure and function of landscape and environments in which they occur [3,4]. They directly and indirectly affect other animal species throughout the food web, including their predators and smaller herbivores, and modify abiotic processes involving nutrient cycles, soil properties, fire regimes and primary production [4,46]. Large carnivores and megacarnivores can exert ecological effects despite existing at low densities [2]. Classically, the effects of large carnivores and megacarnivores were thought to extend down the food web to herbivores and to plants [2]. However, it has been shown more recently that their cascading influences can also propagate broadly to other species mediated by their controlling effects on mesocarnivores, as well as cause subsequent cascades that can affect yet other species [2,4].

Table 1.

Functional traits of focal species considered for trophic rewilding. ABM is average body mass in kg [23]; feeding type is ‘B’ for browser, ‘C’ for carnivore and ‘G’ for grazer; habitats are derived from IUCN Red List species accounts [24]; elevation in m is from the IUCN Red List species account where reported [24], except for the horse as the elevation limit appears to be for its current indigenous range only due to the restricted elevational range provided being between 1100 and 2000 m [24], relative to the known feral range upper limit of 4500 m [25]; conservation translocation type is broken into reintroduction: ‘HIS’ for historical reintroduction and ‘PLE’ for Pleistocene/Prehistoric reintroduction; and ‘ECOL-R’ for ecological replacement. Note that the proposed translocation type depends on area, so the focal species can fall into more than one category. Within each feeding type (carnivore and herbivore: grazing/browser), species are ordered by ABM. See the electronic supplementary material for discussion of each of these species in the published literature.

taxon	common name	ABM	feeding type	habitat	elevation	translocation type proposed	known or potential ecological functions
Panthera tigris	tiger	162	C	mixed grasslands, woodlands, forests	4500	Asia: HIS, PLE [26]	apex predator of large herbivores [26,27]; regulates prey populations [27]
Panthera leo	lion	159	C	grasslands, savannahs, forests, deserts	4200	Africa [27], Asia [28], Europe [29]: HIS, PLE; North America: ECOL-R [30,31]	apex predator of large herbivores [27,30,31]; regulates prey populations [30]
Panthera pardus	leopard	52	C	grasslands, deserts, mountains, forests	5200	Africa [32], Asia [32], Europe [29]: HIS, PLE	predator of large herbivores [29]; regulates prey populations [27]
Acinonyx jubatus	cheetah	51	C	grasslands, dry forests, deserts	4000	Africa [33], Asia [33]: HIS, PLE; North America [30,31]: ECOL-R	predator of fast, agile prey [27,30,31]
Loxodonta africana	African savanna elephant	3825	G/B	grasslands, savannahs, forests, deserts	n.a.	Africa: HIS, PLE; Australia, South America: ECOL-R	ecological engineer by dispersing large seeds and removing trees [1,31,35]; controls plants adapted to grazing [34]
Elephas maximus	Asian elephant	3270	G/B	forests, mixed grasslands	3000	Asia [26]: HIS, PLE; Australia [34], Europe [29], North America [30,31], South America [35]: ECOL-R	ecological engineer by dispersing large seeds and removing trees [1,29–31,35]
Ceratotherium simum	white rhino	2285	G	grasslands, savannahs, shrublands	n.a.	Africa: HIS, PLE; Australia [34]: ECOL-R	ecological engineer by dispersing large seeds and removing trees [30,31]; controls plants adapted to grazing [34]
Rhinoceros unicornis	Indian rhino	1844	G	grasslands, savannahs, shrublands	n.a.	Asia [26]: HIS, PLE; Australia [34]: ECOL-R	ecological engineer by dispersing large seeds and removing trees [26]
Diceros bicornis	black rhino	996	B	shrublands, grasslands, savannahs, deserts	n.a.	Africa: HIS, PLE; Australia [34]: ECOL-R	ecological engineer by dispersing large seeds and removing trees [30,31]
Bison bonasus	European bison or wisent	676	G/B	grasslands, open forests	2100	Europe [36]: HIS, PLE	wallowing creates ephemeral pools, serve as fire breaks, and increase landscape-scale plant diversity [29,36,37]
Camelus dromedarius	dromedary camel	493	B	desert scrub		Africa [27], Asia [28]: HIS, PLE; North America [30,31] ECOL-R	salt-tolerant [29,38]; large home ranges [39], thus may redistribute salts [40]; seed dispersal [31]
Equus ferus	horse	404	G	grasslands, open forests	4500 [25]	Asia, Europe, South America: [41] HIS, PLE, ECOL-R; North America: HIS, PLE, ECOL-R	feeds on coarse, abrasive grasses [41]
Ovibos moschatus	muskox	313	G	Arctic tundra	n.a.	Asia [29], Europe [29] HIS, PLE	grazer in extreme arctic environment [29,37]
Cervus elaphus	red deer	241	G/B	woodlands, shrublands, grasslands, mountains	4500	Europe: HIS, PLE	grazing modifies habitat and provides prey base for carnivores and scavengers [42]
Tapirus terrestris	lowland tapir	169	G/B	forests, shrublands, grasslands	n.a.	North America, South America: HIS [43,44]	disperse seed and browse [26,43,44]
Equus asinus	donkey	165	G/B	deserts	n.a.	North America: HIS, PLE, ECOL-R [31,35]	digs wells used by other species [29,45]

To define the types of translocations for trophic rewilding discussed in the literature, we follow the International Union for Conservation of Nature (IUCN) [47] guidelines for reintroduction and other conservation translocations. The guidelines define conservation translocations as the human-mediated movement and free-release of living organisms intended to yield a measurable conservation benefit at the levels of a population, species or ecosystem, and not only provide benefit to translocated individuals [47]. Conservation translocations are further classified by their primary objective, be it the restoration of declining or extirpated populations (population restoration), or the movement and release of an organism outside its indigenous range (conservation introduction). Population restorations can occur where the species is still present (reinforcement) or extirpated (reintroduction). Conservation introductions can occur to prevent the extinction of the focal species (assisted colonization) or to replace the ecological function of a different, extinct species (ecological replacement). We further classify reintroductions into species extirpated from a region within the last 5000 years ago (historic reintroductions) and more than 5000 years ago (prehistoric reintroductions) [1]. The evaluation of these trophic rewilding candidates and the types of conservation translocations is not intended to be definitive; we use the case studies to test if SDMs can identify climatically suitable habitat now and into the future. We note that if suitable climates are identified by SDMs a more detailed evaluation is needed before trophic rewilding projects proceed, including assessing societal risks, such as threats of megafauna to humans, and ecological risks, such as space and food availability for megafauna.

4. Modelling suitable habitat for trophic rewilding candidates

Here, we used correlative SDMs to model the potential distribution of the focal species (table 1) under current and future climates. To overcome biases introduced by contracted modern ranges [4,48], we generated pseudo-presence locations from estimated present-natural distributions [49]. The term ‘present-natural’ refers to the state that a phenomenon would be in today in the complete absence of human influence through time [50]. Present-natural species ranges therefore refer to the estimated distributions for the present day (notably current climate) in the complete absence of past and present human impacts. The use of pseudo-presence locations in correlative SDMs has already been applied as a way to account for anthropogenic influences [16,17], and has been shown to produce reliable results [16]. We generated pseudo-presence locations at 50, 100 and 200 km for the focal species, excluding locations that were above the highest elevation reported from the IUCN Red List or, if not stated, 4000 m (figure 1 and table 1; [24]). For the horse (Equus ferus), we did not use the IUCN Red List elevation limit, as it appears to be for its current indigenous range only (1100–2000 m [24]), and instead used the upper elevation from the known feral range that goes up to 4500 m [25].

Figure 1.

Present-natural distributions (blue; [49]) and International Union for Conservation of Nature (IUCN) range maps (red lines) for trophic rewilding candidates. The term ‘present-natural’ refers to the state that a phenomenon would be in today in the complete absence of human influence through time [50]. For IUCN range maps, African rhino species are shown at the country level and there are no range maps for the donkey and dromedary camel. Key locations mentioned in the text are marked: IN, introduced range; and CI, conservation introduction.

We modelled climatic suitability for the focal rewilding candidates using the maximum entropy (v. 3.3.3 k; Maxent [51]) and Bioclim [52] predictive algorithms. Full details on methods of SDM construction (including data processing and model evaluation) are provided in the electronic supplementary material. Briefly, we used the two algorithms to ensure robustness in our results. For Maxent models, we conducted species-specific tuning of regularization multipliers and feature classes to balance model fit and predictive ability [53]; Bioclim is less prone to overfitting. Temperature and precipitation are often seen as the main factors driving species distributions at coarse spatial scales [54]. For both algorithms, we used four climate variables: maximum temperature (temperature of the warmest month); temperature seasonality (the difference between temperature of the warmest and coldest month); minimum precipitation (precipitation of the driest month); and precipitation seasonality (the difference between the precipitation of the wettest and driest month). The variables were chosen because they are probably biologically meaningful, because they are weakly correlated at the global scale and because they potentially represent environmental characteristics that limit distributions (see electronic supplementary material for more details). These variables came from the Worldclim database [55] at a 2.5 arc-min resolution, for current climate (average for 1950–2000) and projections for 2070 climates (average for 2061–2080; hereafter future climates). Note that the definitions of temperature seasonality and precipitation seasonality used here are different from the standard versions from Worldclim, although they are directly calculable as the differences between standard Worldclim variables. To account for the uncertainties associated with climate modelling, we use the same variables projected via two general circulation models (GCM). Specifically, we used the GCMs CCSM4 from the Community Climate System Model and CSIRO-Mk3 from the Commonwealth Scientific and Industrial Research Organisation under two representative concentration pathways (RCP2.6 and RCP8.5). The RCPs were chosen because they represent the extremes of likely future greenhouse gas emissions, and thus are representative of the range of possible future climates depending on future social and economic development. RCP2.6 corresponds to a world with mitigation scenarios aiming to reduce emissions; whereas, RCP8.5 is broadly representative of a business-as-usual scenario in terms of patterns of energy usage. In the main text, we focused on the worst-case scenario (RCP8.5); there is little evidence for a sharp decline in fossil-fuel use, which would be needed for the less severe option, and the business-as-usual scenario may become increasingly likely. In addition, we focused on this scenario to estimate the realistically largest difference between current and future climates. To threshold the probabilistic suitability maps we chose the 10% training threshold rule which rejects the lowest 10% of predicted values (OR₁₀), a commonly used approach that reflects the prevalence of species reasonably well [56]. The use of OR₁₀ is likely appropriate when occurrence records have some errors, such as the use of pseudo-presence locations. We restricted SDM projections to the range of environmental conditions corresponding to model calibration by excluding areas estimated from multivariate environmental similarity surface (MESS) maps [57]. For each species, we estimated the relative contribution of each variable to the selected Maxent model based on permutation importance [51]. To measure performance of SDMs, we used the area-under-the-curve of the receiver operating plot (AUC) [58] and for Maxent models we assessed overfitting through threshold-dependent test for the point omission rate based on minimum training presence value (OR_MTP) [58]. We conducted the SDM analyses using the Behrmann cylindrical equal-area projection. Note that exploratory analyses of Maxent and Bioclim models built with pseudo-presence locations separated by 50, 100 and 200 km and Maxent models built with buffers of 1000, 1500 and 2000 km produced similar results, so we present outputs from SDMs for pseudo-presence locations 100 km and for Maxent 1500 km only. All analyses were performed in R v 3.3.1 [59].

5. Results

Across the focal species, SDMs had poor to excellent predictive accuracy; for Maxent models the AUC values were a median of 0.69 and range of 0.62–0.83 and for Bioclim models the AUC values were a median of 0.74 and a range of 0.60–0.90 (electronic supplementary material, table S1). Maxent models were generally not overfit as indicated by low OR_MTP values with a median of 0.07 and a range of 0.02–0.18. Among the climatic variables, temperature seasonality made the greatest contribution for six species, maximum temperature for five species, maximum precipitation for four species and precipitation seasonality for two species (electronic supplementary material, table S1). Predictions of climatic suitability under current climates for the focal species mostly covered the present-natural range, i.e. the prehistoric range of the species, where population restorations (reinforcements and historic- and prehistoric-reinforcements) have been proposed (figures 1 and 2 and electronic supplementary material, figures S1–S34). For suggested conservation introductions, regions of climatic suitability were predicted for both elephant species in most of Australia, parts of southern Europe, and the Americas except for Patagonia and Chile, as well as the Great Plains and northward. For three rhino species, climatically suitable habitat was identified for Australia: most of continent was suitable for black rhinos and most of northern Australia was suitable for white and Indian rhinos. For the three domesticated species (donkey, dromedary camel and the horse) that have established feral populations, climatic suitability was reasonably accurate for their introduced ranges; for example, suitable climate was predicted in Australia for the three species and for much of the Americas for horse and to a lesser extent for donkeys (figure 2 and electronic supplementary material figures S13–S14, S17–S18 and S25–S26). For the donkey, however, the SDMs under-predicted climatically suitable habitat for Central America. For lions and cheetahs, climatically suitable habitat was identified in North America except for north of the Great Plains (figure 2 and electronic supplementary material, figures S29–S30 and S33–S34).

Figure 2.

Current projections of trophic rewilding candidates as modelled by the maximum entropy (Maxent) predictive algorithm. Areas below the 10% training threshold and outside the range of model calibration as estimated by Multivariate Environmental Similarity Surfaces (MESS) maps were clipped from the climatic suitability maps. Each panel has its own colour reference, with warmer colours indicating higher climatic suitability. See electronic supplementary material for high-resolution versions of these maps, as well as maps modelled with the Bioclim predictive algorithm. The maps do not address whether a given area is an appropriate recipient area for a given species in terms of species composition, non-climatic environmental factors or societal acceptability.

For future-climate scenarios, regions proposed for population restoration for the focal species mostly remained or slightly decreased in area of climatic suitability (figure 3 and electronic supplementary material figures S1–S34). However, for some species climatic suitability in the current or former range was projected to decrease substantially, including for the Indian rhinoceros, muskox and the lowland tapir. The decrease in climatically suitable habitat was, in general, projected to be more extensive under the RCP8.5 emission scenario, the representative business-as-usual scenario, compared to RCP2.6, the mitigation scenario that aims to reduce emissions (electronic supplementary material, figures S1–S34). For proposed conservation introduction sites, climate remained suitable for elephants in most of Australia, and expanded northward in Europe and North America (figure 3 and electronic supplementary material, figures S1–S4). Climatic suitability for rhino species remained high in eastern and northern areas of Australia (figure 3 and electronic supplementary material, figures S5–S10). For lions, cheetahs, dromedary camel and donkey, climatically suitable habitat expanded northward in North America (figure 3 and electronic supplementary material, figures S13–S14, S25–S26, S29–S30 and S33–S34). For leopards and tigers, projections of climatic suitability expanded north and east towards Europe for future climates (figure 3 and electronic supplementary material, figures S27–S28 and S31–S32).

Figure 3.

Future projections of trophic rewilding candidates as modelled by the maximum entropy (Maxent) predictive algorithm for the year 2070 via the CCSM4 general circulation model (GCM) and the 8.5 representative concentration pathway (RCP). Areas below the 10% training threshold and outside the range of model calibration as estimated by Multivariate Environmental Similarity Surfaces (MESS) maps were clipped from the climatic suitability maps. Each panel has its own colour reference, with warmer colours indicating higher climatic suitability. See electronic supplementary material for high-resolution versions of these maps and for the CSIRO-Mk3 GCM and RCPs 2.6 and 8.5, as well as the maps modelled with the Bioclim predictive algorithm. The maps do not address whether a given area is an appropriate recipient area for a given species in terms of species composition, non-climatic environmental factors or societal acceptability.

6. Identification of climatically suitable habitat for trophic rewilding candidates using species distribution models

Our SDMs identified climatically suitable areas for widely considered trophic rewilding candidates in regions that have been discussed in the literature, under current and future climates. We were able to identify potential locations that are climatically suitable for future conservation translocations of these focal species, for both population restorations (reinforcements and historic- and prehistoric-reinforcements) and conservation introductions (ecological replacements). Below, we discuss the SDMs for the focal species under current and future climates, and the implications of our findings for trophic rewilding projects under global change.

A key aspect in planning for trophic rewilding projects is identification of suitable sites that match the needs of the focal species into the future [1]. For population restorations, our results broadly suggest large areas of climatically suitable habitat for the focal species under current and future climates (table 1 and electronic supplementary material). Although our findings are not surprising for current climate, given we trained SDMs from pseudo-presence locations generated from presence-natural distributions [49], this provides reassurance for our methodology. Importantly, for future climates the climatic suitability of habitats remained similar or only slightly decreased from current conditions for the majority of species in their former ranges, even under the more severe business-as-usual climate scenario (RCP8.5), although there were more substantial range contractions for some species. For the selected megaherbivores, SDMs projected large areas of climatically suitable habitat under future climates: still, for African savannah elephants and Asian elephants, suitability reduced slightly in northern and central areas of their former range, respectively; for white and black rhinos, suitability decreased in northern parts of their former ranges; and for Indian rhinos, there were extensive contractions in suitability projected, particularly in western and southern areas of their former range (figure 1 and electronic supplementary material, figures S1–S10). Projections of the amount of climatically suitable habitat for large herbivores varied for suggested population restorations: for European bison, dromedary camel, horse and red deer suitability remained similar; for moose and donkey suitability remained similar or slightly decreased; and for muskox and lowland tapirs suitability decreased (figure 2 cf. 3 and electronic supplementary material, figures S11–S26). For the megacarnivores and large carnivores, SDMs projected of climatic suitability remained similar or slightly decreased: for tigers suitability decreased on the Indian subcontinent; for lions suitability slightly decreased in Northern Africa; for leopards and cheetahs, projections suggested similar habitat, although the results depended on the predictive algorithm (figure 2 cf. 3 and electronic supplementary material, figures S27–S32). Nevertheless, in the majority of cases for these focal species, areas in which they are currently found in native ranges remained climatically suitable, with the main exceptions in some areas for muskox, lowland tapirs, tiger and leopards, depending on the predictive algorithm and the climate scenario (electronic supplementary material, figures S19–S20, S23–S24, S27–S28 and S31–S32). At least in terms of climatic suitability, our results suggest opportunities to restore populations throughout their former range for many of these species.

Because of uncertainties around moving species outside their indigenous range, SDMs can be used to determine species' future habitat suitability to guide conservation introductions [60]. We show for conservation introductions, an inherently more risky approach than traditional conservation translocations such as reintroductions [47], climatically suitable habitat for the focal species exist widely in regions discussed in the trophic rewilding literature, now and into the future (table 1 and electronic supplementary material). Among megaherbivores, elephants have been widely discussed as a trophic rewilding candidate across much of the globe [1,30,31]. We identified climatically suitable habitat for elephants under current climates throughout much of Australia, southern Europe, most of the Americas except for Patagonia and Chile, as well as the Great Plains and northward. This identified habitat should remain suitable under future climates for both elephant species, as well as expanding northward in Europe and North America, potentially increasing opportunities for rewilding projects (e.g. [31]). For rhino species, mentioned as possible rewilding candidates in Australia [34], climatically suitable habitat was identified in northern and eastern areas as well as along parts of the southern coast under current and future climates. This, reassuringly, includes potential areas suggested for the establishment of breeding populations of rhinos as insurance populations from poaching in their native range (http://theaustralianrhinoproject.org/index.php). For the large herbivores discussed for conservation introductions including in the Americas [30,31,35], climatically suitable habitat was identified for dromedary camel, donkeys and horses in North America under current climates, and was projected to expand northward under climate change. Climatic suitability for horses in South America remained high for future climates. Among the megacarnivores and large carnivores [30,31], discussed for possible conservation introduction to North America, suitable climates increased for both lions and cheetahs under climate change. Our results, therefore, suggest for many of these focal species that under climate change conservation introductions could occur to areas that should remain climatically suitable until towards the end of the twenty-first century, even under a business-as-usual climate scenario (RCP 8.5).

7. Considerations for the use of species distribution models to determine trophic rewilding opportunities

While there are many considerations that need to be made when building and using SDMs [11,12], the type of SDM approach used to determine trophic rewilding opportunities is conditional on the availability of suitable data and resources [10,13,14]. Although correlative SDMs are most commonly used to model the potential distribution of species under climate change [10,18], methodological factors, such as biased data, can potentially affect their predictions [11]. In this study, the focal species have all experienced anthropogenic range contractions, thus potentially preventing correlative SDMs from accurately projecting climate change effects on their distributions if only current occurrence records were used [48]. This is because assessments of climate niches based on current distributions are likely to be truncated, i.e. not capturing the full realized niche [61], potentially leading to biases in projections under climate change [48]. Although incorporation of past occurrence records, such as from palaeoecological databases, can reduce anthropogenic biases in correlative SDMs [62], this is not always possible when fossil records are incomplete or themselves biased. To overcome these biases, we applied an alternative method by generating pseudo-presence locations [49] from estimated present-natural distributions, which are increasingly becoming available (e.g. [49]). Although pseudo-presence locations have been used to model climatically suitable habitat for three rewilding candidates (Asian elephant, cheetah and lion) in North America under current climates, this previous study generated locations from modern and historical range maps only [63]. Not surprisingly, our results identified climatic suitability for these species over a greater extent of North America as all human-induced range contractions were attempted to be taken into account. Recent studies have advocated for the use of pseudo-presence locations generated from expert range maps when known biases exist [16], including for modelling species distribution under climate change [17,64].

An alternative method that could be used to identify climatically suitable habitat for trophic rewilding candidates is mechanistic SDMs. However, although mechanistic SDMs have great potential to identify suitable habitat under global change and to evaluate the potential effectiveness of management interventions, such as for trophic rewilding projects, they are currently limited to only a few species for which physiological and/or demographic data are available [12–14]. Although the majority of mechanistic SDMs have, to date, been for small-bodied ectotherms [12,13], increasingly endotherms are being modelled with these types of approaches [22,65]. The use of mechanistic SDMs for large-bodied endothermic organisms is particularly challenging [13], such as for the focal species in this study, with some authors recently questioning the assumptions made in some of the more physiologically-informed models [66].

8. Concluding remarks

Trophic rewilding, the (re)introduction of a species into an ecosystem to promote self-regulating biodiverse ecosystems, inherently involves uncertainty [1]. SDMs may reduce uncertainty by identifying climatically suitable habitat for focal species of trophic rewilding projects under current and future climates. We suggest for these trophic rewilding candidates that they will have wide areas of suitable climate available in the regions discussed for their (re)introduction in the literature. Forecasts of suitable climate areas of these species under future climate scenarios suggest that the discussed (re)introductions are likely to remain widely suitable across the twenty-first century. We conclude that SDMs are a powerful tool to identify a species' future habitat suitability and should be used to determine opportunities for trophic rewilding under global change.

Data accessibility

All the data and software used are freely available online. Present-natural range maps are available from https://github.com/MegaPast2Future/PHYLACINE_1.2.

Authors' contributions

S.J. and J.-C.S. developed the ideas and wrote the paper; S.J. performed the analyses, with advice from J.-C.S.

Competing interests

We declare we have no competing interests.

Funding

This study was funded by a Carlsberg Foundation Semper Ardens grant to the ‘MegaPast2Future’ project (grant CF16-0005) and a VILLUM Investigator project ‘Biodiversity Dynamics in a Changing World’ from VILLUM FONDEN (grant 16549).