Wildlife corridors:
do they work and who benefits?
(A slightly abridged version of the proposal funded by the UK Natural Environment Research Council and the US National Science Foundation)
Humans have modified over 75% of the global land area, and the resulting habitat loss and degradation is recognized as the principal driver of biodiversity declines [1]. A major consequence of this modification is that large natural habitats have become fragmented into small patches in a matrix of human-modified land such as farms and cities [2]. Species are threatened by loss of their habitat, but also by increasing isolation of the remaining habitat. This isolation is so damaging because most species depend on dispersal to maintain metapopulations (spatially-linked populations) and track changing environments [3]. Increased isolation severely limits the ability of organisms to disperse between habitat patches by increasing the distance they need to traverse through hostile, human-modified land. The resulting dispersal failure leads to declining genetic diversity and local extinctions [4,5].
Wildlife corridors – swathes of natural habitat between otherwise isolated habitat patches (Fig. 1) – provide a promising response to fragmentation. Corridors are used to increase habitat area and link patches by providing habitat though which species can disperse and avoid the hostile matrix. Theoretically, this should: enhance persistence of local populations; increase metapopulation size and their resilience to perturbations; and allow species to shift their ranges in response to climate change and other pressures [6]. As a result, corridor creation is at the core of regional (e.g., New Mexico’s Wildlife Corridors Act 2019), national (e.g., England’s 25-Year Environment Plan [7]) and international (e.g., the UN’s Connectivity Project [8]) biodiversity policies, as well as rewilding strategies [9]. Corridor initiatives involve conserving existing habitat to link patches, as well as creating new habitat with the same aim. There are many such initiatives, e.g., Yellowstone to Yukon (US/Canada), Paseo Pantera (Central America), and Danubeparks (Central Europe). The resulting corridor networks are extremely large-scale, encompassing 100s of km.

However, we lack an understanding of how such corridors operate in the real world. Most research uses model experimental systems with corridors ≤0.5km long and ≤0.4km wide, in settings where patches and corridors are embedded in a relatively benign matrix rather than the cities and farms that generally form the matrix in the real world [10 . While these experiments demonstrate movement of individual animals through corridors, the scales are orders of magnitude smaller than those of corridors in the real world [11]. Dispersal drops off non-linearly (e.g. exponentially) with distance [12], so the effectiveness of a corridor 0.5km long tells us little about how well a corridor of, say, 5km will work. Furthermore, unlike real corridors, the experimental corridors are not crossed by roads or other human land uses. All these factors may diminish corridor effectiveness in the real world. New studies are therefore needed to address the crucial questions: do corridors counter real-world fragmentation; and what corridor characteristics constrain effectiveness? To address these questions, we need fundamental research into the ecology of species’ movement over large-scales and within complex, modified landscapes.
Indeed, there is debate as to whether fragmentation’s impacts on biodiversity are purely due to loss of habitat, or whether changed habitat configuration is also damaging [13,14]. A key issue is to how impacts are measured, with metrics such as species richness, local abundances or individual movements often used [13]. Gene flow is a more meaningful metric which allows robust analysis of isolation effects, and concomitantly of corridor success, because it represents both inter-patch dispersal and successful establishment of dispersers [15].
Species differ in their responses to habitat loss and isolation. Those that can use farms and cities may experience no isolation, others (generalists) may be able to move in the matrix albeit not as well as in natural habitat, while specialists may avoid the matrix completely. Among mammals, more negative fragmentation responses have been reported for species with large body mass, low mobility, forest specialization, and predatory feeding habits [5,16,17]. But patterns differ among studies, making generalisation difficult. Furthermore, species traits associated with greater harm from isolation might differ from those linked to greater benefit from corridors. Indeed, species may use corridors in different ways. “Passage species” perceive the corridor as a connection between habitat patches, while “corridor dwellers” live and reproduce in the corridor. Any benefit from corridors will therefore depend on the species’ traits. For instance, a larger brain mass may indicate behavioural flexibility and ability to use corridors [18], and a higher reproductive rate may enhance ability to establish new populations [19]. If corridors are to aid multiple species, we need to know which traits are associated with the benefits species derive from corridors, a topic that has been virtually unstudied.
To be useful, corridors must allow sufficient dispersal to maintain metapopulations and to enable species to colonise new areas. Designing corridors to achieve these aims must consider not only corridor and species traits, but also where the corridor should be located to maximise benefits. Corridor designs are often informed by models that represent animal dispersal through landscapes. The most commonly-used tools – circuit theory (CT) and least-cost paths (LCPs)20 – represent dispersal as a function of distance travelled and the costs of traversing alternative land uses. Individual-based “stochastic movement simulation” (SMS) represents dispersal in a more complex way, by incorporating species’ perceptual range and behavioural stochasticity. For two datasets, a SMS we developed predicted connectivity (genetic distance) substantially better than LCP or CT21. But, while more complex models may be more accurate, they require estimation of additional parameters. A central question in ecology is whether the accuracy gained is sufficient to justify using complex rather than simple models22, and this issue has not been addressed for connectivity.

Objectives: The central question is: do corridors provide connectivity for wildlife in the real world? Using 20 landscapes in Europe and the Americas, we will deploy a natural experiment approach [23], whereby each landscape contains three types of habitat configuration: a large intact natural area, isolated patches, and patches connected by a corridor, (Fig. 1, 2). Using long-term gene flow as a robust measure of isolation impact and corridor success, and focusing on mammals, which are strongly affected by fragmentation [24], we will answer the following questions.
Q1. What characteristics of corridors determine their success? We will identify the corridor characteristics that most strongly influence success (gene flow) for a range of mammals. We hypothesize a strong non-linear decline in success with increasing corridor length, reflecting the skewed distribution of dispersal distances within species. We also expect success will drop steeply as corridor width falls below a threshold, with the threshold determined by species traits. This will be the first study of the success of large corridors for multiple species, using a robust response variable.
Q2. What types of species benefit most from corridors? We will analyze how species differ in the degree to which corridors enhance gene flow, and relate these to species’ traits. We hypothesize that species that are bigger, are habitat specialists, or have greater dispersal abilities, relative brain size or reproductive rate will benefit more from corridors. This will allow generalization to a wide range of mammal species not covered in this project.
Q3. Which connectivity models best predict corridor success? We hypothesize that more complex models will perform best as they represent better the details of animal ecology, corridor characteristics, and landscape influences. We expect accuracy gains for more complex models to be substantial, allowing us to explore the most effective corridor designs for a range of species types.
Methodology and approach
Natural experiments involve overlaying a robust statistical design on a system that has had manipulations which are out of the researcher’s control, but which can be used as experimental treatments. They fall between true manipulative experiments and observational studies. As they can cover larger scales than the former and are more rigorous than the latter, they have been central to key insights in ecology and other sciences [23]. We will study ~4 focal species in each of ~20 landscapes in Europe, the Americas (Table 1), and perhaps elsewhere. As summarized in Fig. 2, each landscape contains three types of habitat configuration, which will be our treatments: isolated habitat patches, >1 pair of patches connected by a corridor, and a large intact natural area (Fig. 1). The landscapes are ideal [11] replicates for the natural experiment because (1) corridor widths vary from 0.16 to 2.5km and lengths from 2 to 25km; (2) the configurations have been in place for ≥40 years, so genetic patterns will reflect landscape structure; (3) to address recommendations that fragmentation studies should control for habitat area [13], we will match sizes of isolated and connected patches as well as their inter-patch distances; and (4) the focal species span a range of traits. With previous funding, Gregory visited 93 potential landscapes in 37 countries, confirmed the presence of suitable focal species, and met with potential partners, identifying 86 landscapes with de facto corridors. Globally, few corridors have been implemented, all <20 yrs ago, which is too short a time for genetic patterns to reflect the influence of the corridor [25]. Our de facto corridors comprise old, remnant habitat. Importantly, the matrix in each landscape comprises predominantly agricultural and urban land, presenting a strong contrast to the natural habitat areas, which will influence mammal dispersal decisions and behaviors, and the costs incurred. We are selecting landscapes (Table 1) to include even representation of corridor widths, lengths, and road types, to work efficiently close to our home institutions, and to benefit from committed local partners who will provide essential help.
Mammal species. From species suggested by local partners (Table 1), we will select ~4 mammals in each landscape that occur in the focal habitat, avoid the matrix, and represent a variety of traits putatively related to corridor use and benefits. These comprise two orders with quite different characteristics: rodents and carnivores. Although the constrained taxonomy limits generality, it allows inferences to be drawn that are not confounded by large phylogenetic differences. Trait information will be drawn from available databases [e.g., 26-29]. Traits relevant to our hypotheses include body size, diet type, habitat breadth, age at sexual maturity (a measure of reproductive rate), brain mass, and dispersal ability (estimated from other traits [12]).
Measuring effective movement. We will use genetic similarity to measure relative gene flow in each corridor. Although genetic equilibrium probably never fully occurs in any human-altered landscape, genetic differentiation is evident in as few as three generations after perturbations [30], and our habitat configurations have been stable for >20 generations. We will use live traps, hair snares, and faecal sampling to collect tissue samples of each focal species in each habitat patch (Fig.2: A,B,M,N) and in locations in the intact area that are the same distance apart as the patches in the other configurations (Fig.2: Y,Z). We will also sample at key locations in the corridor (e.g. near roads or villages) to identify barriers to gene flow. We will identify samples to species or species group (hair and faeces) to select those for DNA extraction. We will amplify each DNA sample at ≥50 SNPs (single nucleotide polymorphisms) and at 6-10 microsatellite markers (to control for marker bias). We probably need ≈30 individuals per patch which, after post-hoc screening for close relatives, should yield >20 samples. We will take a mobile sequencing lab (MinIon) to each field site, allowing us to ascertain promptly when we have an adequate sample size to detect differentiation, thus optimizing our sampling efficiency. Gregory has extensive experience in all these methods [31]. Libraries of 2-31 SNPs are available for most focal species, and we will develop additional SNPs [32] as needed to reach 50-100 SNPs; the exact number will be based on the power of the selected panel to resolve genetic similarity across the habitat configuration, with an emphasis on surveying more individuals as opposed to more markers [33].
The Corridor Success Index. CSI [15], measures gene flow provided by the corridor for each species in each landscape. The CSI equation (Fig. 2) rescales the genetic similarity between populations in the corridor-connected patches. CSI is the amount by which gene flow through the corridor exceeds gene flow across the matrix, standardised to the gene flow across the intact habitat configuration. A value near 0 indicates the species gains no benefit from the corridor (no greater gene flow than between isolated patches), a value close to 1 shows gene flow in the corridor similar to intact habitat, and intermediate values indicate the degree to which the corridor enhances gene flow. We hypothesize that even for species that are likely to be able to cross the matrix to some extent, the corridor will still enhance connectivity, and so CSI will be positive, albeit near 0.
We will use random forests (RF) models to identify which (Q1) corridor characteristics (length, mean width, road density, years since fragmentation) and (Q2) species traits (body size, diet, etc.) promote CSI. A categorical “landscape” variable will account for dependence among observations within each landscape in a Mixed Effect RF approach [34]. Similarly, a categorical “species” variable will account for non-independence where the same species are sampled in different landscapes. RF does not assume any parametric relationship between response variables and predictors or that residuals are independently and identically distributed. Furthermore, RF’s variable importance (VI) values are robust to multicollinearity and RF can model relationships even when predictors greatly outnumber observations [35]. We will examine interactions among the important predictors using partial dependence plots [36]. In addition, for each corridor/species combination with a low CSI, we will use a Hierarchical Bayesian formulation [37] of boundary detection analysis with BoundarySEER software to identify segments of the corridor with the greatest discontinuity in gene frequencies. Post hoc permutation procedures will measure congruence of these segments with putative barriers such as a road. Because RF does not formally test hypotheses, we will use a 2-stage process to test our hypotheses about the effect of corridor characteristics and species traits, e.g., that large-bodied species benefit more from corridors: (1) Calculate each predictor’s Conditional Variable Importance (CVI)38 (i.e., VI after accounting for multicollinearity), rank predictors in order of CVI, and create reduced RF models by sequential backwards elimination, stopping when mean squared error starts to increase39. If a predictor (e.g., body mass) is not retained in the final model, the hypothesis will be rejected (body mass has no independent influence on CSI). (2) If the predictor is retained, a 1-tailed t-test will determine if, say, large-bodied species have higher CSI than small-bodied ones.
Table 1. Landscapes and Focal Species. Habitat configuration has been stable ≥20 generations of the focal species. In each landscape, the matrix is dominated by human land uses. Width is at the narrowest portion of each corridor; mean width of each corridor is 50% to 200% wider.

Comparing connectivity models. We will compare the performance of a range of connectivity models (Q3): (a) the simple straight-line Euclidean distances among locations; (b) the widely used Least Cost ath (LCP); (c) cumulative pairwise resistance distances (derived from Circuit Theory, CT); (d) pairwise dispersal rates simulated with the SMS; and (e) pairwise dispersal rates estimated through RangeShifter (RS), a complex individual-based model. We will do this for each of our focal species in each landscape. This will extend an “inverse fitting” approach we have tested previously by applying Approximate Bayesian Computation (ABC) to estimate the parameter values for each model that give best fit to the genetic data21. We expect the best fit versions of each model (a)-(e) to vary in their goodness of fit and we will use this to compare the accuracy of each model and thus whether more complex models give significantly better fit. It is relatively straightforward to use ABC for Euclidean, LCP and CT connectivity metrics, and we have developed versions of SMS and RS that are integrated with ABC. We recently tested our SMS-ABC model by estimating parameters for the Cabanis’s Greenbull, which produced results that matched detailed movement data very well [40].
How much model complexity is needed for corridor planning? Even if more complex models (e.g. SMS or RS) fit the spatial genetic data better than simpler models, this does not necessarily mean that a more complex model is needed to inform corridor design (Q3). Different models might suggest the same corridor design. Comparing corridor designs is the key test of whether it is worthwhile to use more complex modelling approaches, but our project will be the first to conduct such a comparison. Alternative landscape designs, based on practicality and local plans will be drawn up by working with our partners and key stakeholders in at least 2 European and at least 2 American landscapes. We will use each connectivity model (LCP, CS, SMS and RS) to rank these alternatives. The rankings will be conducted at different levels: (1) by using the best fitting LCP, CS, SMS and RS models for each species in turn; (2) aggregated across all species, by generating a mean rank across species, and also weighted means whereby key species (identified by partners) are given extra significance. We will compare the rankings from the different models and the degree to which they concur or differ22. We will also determine whether models yield similar management recommendations for all species, or whether they differ in relation to species traits; e.g. if smaller-brained species need shorter and wider corridors. From our results, we will identify the conditions (if any) that favour investment in more complex modelling and in research to estimate the necessary parameters.
Scientific excellence. The sampling of genetic structure across a range of species in 20 landscapes will provide unprecedented insights into how mammals respond to changed connectivity in the real world. Working closely within a team with expertise in field ecology, genetic analysis, statistics and modelling, we will assess how long, wide and well-protected a corridor must be to aid biodiversity recovery (Q1), which types of species benefit most (Q2), and which models are best for designing effective corridors (Q3). While such information is important and timely for making connectivity plans scientifically rigorous, the project will also provide novel insights into species dynamics at large scales, which will transform spatial population ecology. We will, for the first time, measure the effective movement of multiple species at scales of a few to 10s of km (Table 1); that is, movement which leads to the establishment of populations and thus gene flow. We will be able to assess how this movement is modified by landscape configuration; not only inter-patch distances and dispersal routes provided by corridors, but also how human land uses, such as roads, alter effective movement. We will, furthermore, provide detailed understanding of how and why mammal species vary in their effective movement in response to fragmentation and landscape configuration. By linking our gene flow data to the ever improving data on mammal traits, we will ascertain if and how dispersal ability, habitat specialisation, behavioural flexibility or population growth rate characterise differences in effective movement. Finally, we will, for the first time, use robust field data to contrast different ways of representing dispersal and demography over large scales in ecological connectivity models. This will allow generalisation about species’ responses to changed connectivity, and provide fundamental insights into how model complexity and accuracy inter-relate. These insights will lead to papers in high impact journals, and essential guidance for stakeholders on corridor design.
Literature Cited:
[1] IPBES. The IPBES assessment report on land degradation and restoration (2018)
[2] Haddad et al. Science Advances 1 (2015)
[3] Cosgrove et al. Divers Distrib 24, 448 (2018)
[4] Bullock et al. Trends Ecol. Evol. 33, 958 (2018)
[5] Lino et al. Mamm. Biol. 94, 69 (2019)
[6] Isaac et al. J. Appl. Ecol. 55, 2537 (2018)
[7] Defra. A Green Future: Our 25 Year Plan to Improve the Environment, https://www.gov.uk/government/uploads/system/uploads/attachment_data/file/693158/25-year-environment-plan.pdf (2018)
[8] UNEP. Connectivity Conservation, https://www.protectedplanet.net/c/connectivity-conservation/overview-of-unep-led-initiative (2018)
[9] Perino et al. Science 364, 5570 (2019)
[10] Gilbert-Norton et al. Conserv. Biol. 24, 660 (2010)
[11] Beier & Gregory. PLoS Biol. 10, e1001253 (2012)
[12] Santini et al. Glob. Chang. Biol. 22, 2415 (2016)
[13] Fahrig et al. Biol. Conserv. 230, 179 (2019)
[14] Fletcher et al. Biol. Conserv. 226, 9 (2018)
[15] Gregory & Beier. Conserv. Biol. 28, 689 (2014)
[16] Vetter et al. Ecography 34, 1 (2011)
[17] Keinath et al. Glob. Ecol. Biogeogr. 26, 115 (2017)
[18] Santini et al. Ecol. Lett. 22, 365 (2019)
[19] Sibly & Brown. Am Nat 173, 185 (2009)
[20] Benz et al. PLoS One 11, e0162989 (2016)
[21] Coulon et al. Ecology 96, 2203 (2015); 22Willcock et al. Ecosystems (2019)
[23] Watts et al. Ecol. Evol. 6, 3012 (2016)
[24] Crooks et al. Proc. Natl. Acad. Sci. U. S. A. 114, 7635 (2017)
[25] Beier et al. Conserv. Biol. 22, 836 (2008)
[26] Wilman et al. Ecology 95, 2027 (2014)
[27] Jones et al. Ecology 90, 2648 (2009)
[28] Rondinini et al. Philos. Trans. R. Soc. Lond. B Biol. Sci. 366, 2633 (2011)
[29] Gonzalez-Voyer et al. Evolution 70 (2016)
[30] Landguth et al. Mol. Ecol. 19, 4179 (2010)
[31] Winder et al. Ecosphere 6, 17 (2015)
[32] Aitken et al. Mol. Ecol. 13, 1423 (2004)
[33] Fumagalli et al. Genetics 195, 979 (2013)
[34] Stephan et al. Nature Comm 6 (2015)
[35] Witten et al. An introduction to statistical learning. Springer (2013)
[36] Cutler et al. Ecology 88, 2783 (2007)
[37] Lu & Carlin. Geogr Anal 37, 265 (2005)
[38] Strobl et al. BMC Bioinformatics 9, 307 (2008)
[39] Murphy et al. Ecology 91, 252 (2010)
[40] Aben et al. J. Appl. Ecol. 53, 1055 (2016)
[41] Bocedi et al. Methods Ecol. Evol. 5, 388 (2014)
[42] Synes et al. Ecol. Inform. 30, 222 (2015).