Evaluation of camera placement for detection of free-ranging carnivores; implications for assessing population changes

1 Centre for Integrative Ecology, School of Life and Environmental Sciences, Deakin University, Burwood, Australia 2 Threatened Species Recovery Hub, National Environmental Science Program, Research Institute for the Environment and Livelihoods, Charles Darwin University, Darwin, Australia 3 Parks Victoria, Halls Gap, Australia 4 Institute for Land,Water and Society, School of Environmental Sciences, Charles Sturt University, Albury, Australia 5 Phillip Island Nature Parks, Cowes, Australia


INTRODUCTION
Introduced mammalian carnivores are among the greatest threats to biodiversity (Doherty, Glen, Nimmo, Ritchie, & Dickman, 2016), having been associated with the decline and extinction of numerous species worldwide. In Australia, feral cats Felis catus and red foxes Vulpes vulpes have contributed to declines in many native bird and reptile populations (Doherty et al., 2016), and together have been implicated in most of the thirty mammal extinctions that have occurred since European settlement (Woinarski, Burbidge, & Harrison, 2015). This has triggered the development of various management approaches aimed at mitigating their impacts, such as predator free fencing (Legge et al., 2018), translocation to predator free islands (Abbott, 2000), guardian animals (van Bommel, 2010;van Bommel & Johnson, 2012) and lethal control (Doherty & Ritchie 2017;Doherty, Driscoll, Nimmo, Ritchie, & Spencer, 2019;Hunter, Lagisz, Leo, Nakagawa, & Letnic, 2018;Molsher, Newsome, Newsome, & Dickman, 2017). Each of these management approaches vary considerably in cost, spatial extent and effectiveness.
It is therefore vital to monitor introduced carnivore populations to ensure management efforts are achieving their intended outcomes.
It can be challenging to evaluate whether management interventions are working. This is typically because cats and foxes are cryptic, they tend to occupy relatively large home-ranges, and sometimes occur in low densities (Balme, Hunter, & Slotow, 2009), leading to low detection probabilities and difficulties associated with developing feasible monitoring programmes. Several studies have shown a positive association between introduced carnivores and open or fragmented habitats (e.g. forest edges, recently burnt areas, Graham et al., 2012), which may be due to an increase in prey vulnerability or density (Hradsky et al., 2017). Roads have been shown to facilitate carnivore movement to these sites (Hradsky et al., 2017), and consequently, the proximity of traps to roads and other modified features are often considered as part of targeted carnivore surveys. Indeed, many studies have focused their efforts on roads due to perceived increases in detectability (Bubela, Bartell, & Muller, 1998;McGregor, Legge, Potts, Jones, & Johnson, 2015;Towerton, Kavanagh, Penman, & Dickman, 2016).
Another challenge is that co-occurring species often interact, either through direct effects (i.e. interference competition and intraguild predation) or indirect effects (i.e. fear-mediated behavioural change) (Ritchie & Johnson, 2009). This is likely to affect the efficacy of multispecies monitoring programmes. For example, Hayward and Marlow (2014) suggest that subordinate carnivores avoid roads in areas where they co-occur with a more dominant (apex) carnivore. By contrast, other studies have found that both subordinate and dominant carnivores select for roads, using them frequently, and sometimes simultaneously (Mahon, Bates, & Dickman, 1998;Read & Eldridge 2010;Wysong, Iacona, Valentine, Morris, & Ritchie, 2020).
Feral cats generally occupy a mesopredator role and may be suppressed, to varying extents, by larger carnivores. For example, in parts of Australia, dingoes Canis dingo, Tasmanian devils Sarcophilus harrisii and foxes have been shown to suppress cats (Brook, Johnson, & Ritchie, 2012;Cunningham, Johnson, & Jones, 2020;Marlow et al., 2015), possibly leading to altered behaviour (Molsher et al., 2017). If this is the case, we might expect cats to be detected less in areas where larger carnivores are most active in space and time. For example, if foxes are using roads, cats might avoid them (i.e. spatial avoidance)or use them at different times (i.e. temporal avoidance) -to reduce the probability of an encounter. Spatial avoidance behaviour would have implications for how we monitor co-occurring carnivores (Hayward & Marlow, 2014), as it means the optimal approach for monitoring a dominant carnivore would differ from that of subordinate carnivores, potentially requiring different monitoring approaches for each. By contrast, if such avoidance behaviour is not evident, or if avoidance is largely temporal rather than spatial, then a single broad approach (e.g. monitoring on roads) might adequately capture both types of carnivores, noting that the specific design may require some optimization to deal with differences in density or rates of change, which may vary between species (regardless of whether they avoid one another or not).

F I G U R E 1
The location of the Grampians National Park relative to Australia (a) and the approximate location of paired camera sites within the Grampians National Park (b) In this study, we assessed our ability to effectively monitor cats and foxes in the Grampians National Park (GNP), an area of high biodiversity and conservation value in central-west Victoria, Australia. We used motion sensor cameras to test whether cat and fox detectability was greater on-road compared with off-road habitats, with explicit consideration of how this affects our ability to detect changes in occupancy. We also assessed whether there was any evidence of spatial or temporal interaction between cats and foxes that may influence their road-use. While we acknowledge that red foxes are native in a substantial proportion of their global distribution, our focus here is in the context of invasive species management. Broadly, we seek to inform and aid improvement in the ways these species are surveyed, given the widespread distribution and environmental damage caused by introduced mammalian carnivores globally. Nevertheless, our results may be applicable to other contexts where land managers are interested in monitoring changes in native fox populations, or for other native terrestrial carnivores.

Study location
The GNP encompasses an area of ∼168,000 ha in south-eastern Australia, approximately 260 km west of Melbourne ( Figure 1a)

Survey design
Sites were selected to complement an ongoing, long-term ecological study of small mammals aimed at assessing their responses to wildfire and underlying climatic conditions (see Hale et al., 2016 Molsher et al., 2017;Moseby, Stott, & Crisp, 2009). Neighbouring sites were separated by at least 2 km.
Cameras were mounted to a metal fence post if placed off-road or a security post (within security boxes, design adapted from Meek, Ballard, & Fleming, 2013) if placed on-road (to minimize incidence of theft).
Cameras were positioned facing south, tilted slightly down and 90-100 cm above the ground to ensure a focal point 5-6 m away from the device and to minimize false triggers associated with sun glare (as recommended by Meek, Ballard, & Fleming, 2012 for targeting introduced carnivores). All cameras were set up passively (i.e. no lure), as we were interested in determining the ability of each camera to pick up natural carnivore movements through the landscape, which would be confounded if individuals were lured to cameras. Camera sensitivity was set to high with a quiet period of 5 seconds and a one second delay between images, reflecting the fastest trigger time and lowest delay for this model of camera (Meek et al., 2012). Each event was set to capture three images at a low resolution ( Each three-photo sequence was treated as a single point in time and an event was defined as a set of images separated by 5 minutes -this was considered adequate, as cat and fox resident times (i.e. the amount of time spent within the focal view of the camera) were short, and in most cases limited to a single three-photo sequence. We consider that individuals had no incentive to remain at camera stations and were likely to be passing through.
Images were processed using C PW Photo Warehouse, a custom Microsoft Access application designed to facilitate archiving, identifying, summarizing and analysing photo data collected from remote wildlife cameras (Ivan & Newkirk, 2016).
We sampled cat and fox populations in the GNP across five discrete

Statistical analysis
We used single-season occupancy models (MacKenzie et al., 2002) to estimate occupancy and detection probabilities of cats and foxes in the GNP. We summarized camera observations into 24-hour detection histories, considering each sampling night at each camera one detection attempt. Models are formulated in terms of parameters i and p ij , where i is the probability that site i is occupied by the species of interest and p ij is the probability of detecting the species at site i during survey j, conditional upon it being present. In its basic formulation, the model structure assumes independence among sites and detections, no changes in the occupancy status of sites (i.e. sites are 'closed' -either occupied or empty -across the whole survey period) and no false positive records.
Where sites are spaced too close together with respect to the territorial patterns of the target species (e.g. where the home-range of an individual overlaps with more than one camera), modelled estimates of occupancy and detectability may be biased (MacKenzie & Bailey, 2004).
Given the large home-range sizes of cats and foxes in Australia (Carter et al., 2012;Molsher, Dickman, Newsome, & Müller, 2005), there would be potential for us to violate the assumption of independence of sites if we considered every camera trap a separate site, especially those within a single pair. Therefore, we instead fit three separate models to the data: (i) where detections were pooled across both cameras within a pair (i.e. pooling off-and on-road detections for a given night); (ii) using only off-road detections; and (iii) using only on-road detections.
While sites (i.e. pairs of cameras) were typically spaced >2 km apart, this too may be insufficient to ensure independence: a study conducted elsewhere in southern Australia (within similar habitat) showed that fox home-range sizes were up to 7 km in length (Hradsky et al., 2017).
The placement of cameras on tracks and roads could also increase the risk of non-independence among sites (Hines et al., 2010), given these carnivores have been shown to move large distances along roads over short time periods (Hradsky et al., 2017). No data are available on cat or fox home-range within the GNP, so we tested whether potential violation of this assumption was likely to influence our modelled outputs of occupancy and detectability by fitting models to a subset of data, including only information from on-road cameras spaced >7 km apart (in any straight-line direction).
Similarly, where sites are not closed, modelled outputs may be biased (MacKenzie & Bailey, 2004). However, complete closure is difficult to achieve, particularly so for dynamic environments in continuous habitat, and where the species of interest are mobile (Steenweg, Hebblewhite, Whittington, Lukacs, & McKelvey, 2017). One proposed solution for dealing with potential violations of closure is to redefine the estimated parameter from occupancy (i.e. the probability of occurrence at a given site) to use (i.e. the probability of use of a given site) (Latif, Ellis, & Amundson, 2016). We apply this definition of occupancy here.
Models were fitted within the maximum-likelihood framework for inference using the package 'unmarked' (Chandler et al. 2020) in R (R Core Team, 2019). We did not include any predictors of occupancy or detectability due to small sample sizes and issues associated with model convergence. Additionally, we did not consider multiseason models (MacKenzie, Nichols, Hines, Knutson, & Franklin, 2003) because we were not interested in extinction and colonization dynamics, but rather typical detection probabilities.
Using the fitted detection probabilities obtained from each model, we calculated the probability of detecting each species at site i at least once after k repeat visits as p * = 1 − (1 − p) k , where p * is the cumulative detection probability.

Power analysis
Power analysis allows us to determine whether a given design has the potential to produce a statistically significant result when the effect size (in this case, a change in occupancy) is biologically important.
Guillera-Arroita and Lahoz-Monfort (2012) provide approximations (Equation 1) to calculate how the power of a given study design changes depending on the allocation of survey effort (i.e. number of sites and trap nights).
The probability of observing a significant difference in occupancy (i.e. power), given a significance level of α, is where is the probability of performing a type II error (i.e. not detecting an effect of a given magnitude when one has occurred), 1 and 2 are the true underlying occupancy probabilities in time 1 and 2, Φ(x) is the cumulative distribution function for the standard normal distribution, z ∕2 is the upper 100 /2-percentage point for the standard normal distribution (e.g. 1.96 for = 0.05), 2 We defined R to be the proportional difference in occupancy, so that 2 = 1 (1 − R), with R > 0 representing a decline. For a given R, the power to detect the decline increases both as the number of sampling sites (S) and the number of repeat visits (k) increases.
Here we apply Equation (1) to the fitted estimates of detectability obtained from the occupancy models to test our capacity to detect a decline in occupancy under three hypothetical monitoring regimes.
To test the influence of declining occupancy on our ability to detect a response, we consider three initial starting occupancy probabilities: (i) low (i.e. 1 = 0.3); (ii) moderate (i.e. 1 = 0.5); and (iii) high (i.e. 1 = 0.8). We assume a standard sampling design with k trap nights, across S sites, and fitted pooled, off-and on-road estimates of detectability as averaged across the five sampling occasions; winter 2016, autumn 2017, spring 2017, autumn 2018 and spring 2018. Our calculations assume that two datasets are collected: one at time 1 and one at time 2. The datasets are then analysed and their estimated occupancy probabilities with associated uncertainties compared, to assess whether there is evidence of a decline between the two times considered. We apply this approach to the following three scenarios: 1. Pilot: 30 sites sampled, with two cameras deployed at each site (one on-road and one off-road) for a minimum of 60 nights (as in the present study); 2. Scenario A: 60 sites sampled, with one camera deployed at each site (off-road) for a minimum of 60 nights (effort targeted entirely offroad); and 3. Scenario B: 60 sites sampled, with one camera deployed at each site (on-road) for a minimum of 60 nights (effort targeted entirely onroad).
Our three scenarios considered some of the trade-offs in sampling design for increasing statistical power, specifically by: (i) increasing the number of sample sites (Scenario A and B); and (ii) increasing the number of detectors (i.e. cameras) at a given site to maximize the chance of an encounter (Pilot). We assume a standard sampling duration of 60 nights for all three scenarios because this was sufficient for obtaining high confidence (>95%) that failure to record a cat or fox on a camera reflects a true absence (based on pooled detection histories across onand off-road cameras; see Results), and because longer survey durations are likely to increase the probability of changes to the occupancy status of sites (and thus are more likely to violate the closure assumption). The number of sites sampled was capped based on what could be realistically implemented within the study area based on logistical (e.g. maintaining appropriate spatial distance and replication within the boundaries of the park) and financial (e.g. resources to cover equipment and personnel costs) constraints.
For all of our analyses, we set alpha (α) to 0.05 and beta (β) to 0.95. This assumes equal importance is given to the probability of performing a Type I error (detecting a false decline) as to the probability of performing a Type II error (not detecting a decline when one has occurred) (Di Stefano, 2003).

Temporal interactions
To examine temporal avoidance between cats and foxes, we created temporal activity profiles across each sampling season using timestamps from camera photos ('overlap' package in R; Meredith & Ridout, 2018). This analysis considered only temporal interactions at on-road cameras, due to data limitations (i.e. there were too few off-road detections in some seasons to allow analysis). We plotted the smooth kernel density functions to create a probability density distribution for each species activity pattern and calculated the coefficient of overlap (Δ), which measures the total overlap between the two species temporal activity distributions (ranging from 0 or no overlap to 1 or complete overlap). We calculated 95% confidence intervals using 5000 smoothed bootstrap samples for each species (after adjusting for bootstrap bias; Meredith & Ridout, 2018). We used the estimator Δ 4 for inference, because simulation studies conducted by Ridout and Linkie (2009) and Meredith and Ridout (2018) found that this was the best performing option when the smaller of the two samples was >75.
To further explore the similarity between cat and fox activity patterns, we used Mardia-Watson-Wheeler tests ('Circular' package in R;

Model assumptions
There was little difference between estimates of occupancy and detectability obtained from models fit to all of the available on-road data, compared with models fit to a subset of the on-road data (taken from cameras spaced >7 km apart) (see Supplementary Material S1).
Given that these differences are likely to have a negligible influence on the outcomes of this study, we use the estimates of occupancy and detectability computed using all of the available on-road data for further inference.

Occupancy and detectability
We obtained >60 nights of data from all 30 paired sites (at both camera occupancy was typically lower for cats and foxes compared to pooled and on-road occupancy (with some exceptions); however, the off-road estimates were almost always more imprecise, suggesting a high level of uncertainty in the modelled outputs (particularly so for cats, Figure 2a and 2b).
Detectability varied considerably between on-and off-road locations ( Figure. 2c and 2d), improving significantly for both cats (sevenfold) and foxes (threefold) on-road ( Figure 3). There was almost no difference between the pooled and on-road detectability for cats among seasons (Figure 2c), suggesting that the off-road cameras provided little additional benefit in terms of detecting this species. There were some differences in the pooled and on-road detectability for foxes among seasons ( Figure 2d); however, this was not significant (evident by overlapping confidence intervals). When targeted on-road or when pooled across locations, both predators could be detected with >95% confidence in areas where they were present with <60 trap nights (42 trap nights required for cats and 29-32 trap-nights required for foxes, Table 1). By contrast, 299 trap-nights (cats) and 99 trap-nights (foxes) were required to obtain 95% confidence in absences for cameras deployed off-road (Table 1).
The off-road detectability for cats was unusually high in spring 2018 compared with previous estimates (Figure 2c), despite being recorded only twice at one off-road camera (see Supplementary Material S2).
However, the confidence intervals were wide (ranging from 0.004 to 0.14, Figure 2c). Notably, a greater level of precision (evident by narrower confidence intervals, Figure 2c)

Power analysis
Power to detect declines in occupancy varied among monitoring regimes, species and according to initial starting occupancy probabilities ( 1 ). Scenario B (effort targeted entirely on-road) yielded the greatest power for detecting declines in both cats and foxes under all values of 1 . Scenario A (effort targeted entirely off-road) yielded the least power for detecting declines in cats, however performed better than the Pilot scenario for detecting declines in foxes ( Figure 4).
As 1 increased, our ability to detect a response also increased, leading to greater power for detecting declines of smaller magnitudes.
For example, when 1 = 0.3, only large declines (i.e. >80%) could be detected with ≥95% confidence for both species, and only when survey efforts were targeted entirely on-road (i.e. under Scenario B, Figure 4).
A sampling regime that targets foxes off-road is capable of detecting declines in occupancy, but only of magnitudes greater than 50% (with ≥95% confidence) assuming high 1 (i.e. ≥0.8, Figure 4). No magnitude of decline was detectable for cats with ≥95% confidence when survey effort was targeted entirely off-road (Figure 4).

DISCUSSION
The development and implementation of sound monitoring programmes is integral to cost-efficient and ecologically effective wildlife management and conservation (Robinson et al., 2018). Here we demonstrate that monitoring cat and fox populations using road and track networks in natural landscapes improves our ability to detect both species, leading to improved precision around modelled estimates, increased statistical power and consequently allowing for detection of smaller changes in species occupancy. While we acknowledge that this approach is likely to have some limitations (i.e. limited inference about predator-prey interactions, or carnivore habitat use at off-road sites), we highlight that such large differences in detection rates are likely to have major implications on the quality of data collected, and subsequently, the types of analyses that can be performed. As we show here, data limitations (in this case associated with an off-road approach to monitoring) can lead to inability to perform a given analysis (i.e. temporal activity), uncertainty in modelled estimates (low precision), low power for detecting changes in populations (especially if initial population sizes are small) and potentially poor ecological inference.
Average cat and fox occupancy estimates were relatively high (≥0.53) regardless of the underlying data (i.e. pooled, off-road or onroad), which contrasts with previous estimates from this landscape. Robley et al. (2012) estimated cat occupancy to be approximately 0.17 (SE ± 0.046) in areas of high conservation value, while a broader survey across a larger area estimated fox occupancy to be approximately 0.28 (SE ± 0.086), figures of which are considerably lower than our comparable off-road estimates. While this could be due to genuine differences in occupancy, another explanation is that the cameras in that study were deployed for an insufficient period of time (∼23 days for cats and ∼28 days for foxes) to enable high confidence that these species would be detected if they were present. Indeed, the authors of that report suggest that the cumulative detection probability did not exceed 67% on average (noting that detectability varied according to location, and for foxes was more likely closer to roads). In both cases, although to a lesser extent in Robley et al. (2012) (likely due to a larger sample size), the confidence limits around the occupancy estimates were broad, indicating moderate to high rates of imprecision, and thus should be interpreted with caution.
Our results provide strong support for a positive association between introduced predators and roads, adding to the growing body of evidence that suggests a significant positive effect of roads on predator activity (Carter, Potts, & Roshier, 2019;Dawson et al., 2018;Raiter, Hobbs, Possingham, Valentine, & Prober, 2018;Wysong et al., 2020). The potential for roads to facilitate predator movements has been widely reported in the literature, with several studies documenting the frequent use of roads by predators (Bischof, Gjevestad, Ordiz, Eldegard, & Milleret, 2019;Read et al., 2015) and others deliberately targeting roads to enhance the likelihood of capture (Bubela et al., 1998;McGregor et al., 2015;Towerton et al., 2016).
However, there remains some disagreement about whether cooccurring predators should both use roads preferentially (Haywood & Marlow, 2014;Mahon et al., 1998;Nimmo, Watson, Forsyth, & Bradshaw, 2015;Read & Eldridge, 2010), and this is likely to be context dependent. While we did find some evidence of temporal segregation of cats and foxes (evident in significant differences in peak activity times in all but one season), they overlapped considerably in their activity. This, coupled with high spatial overlap (i.e. both species showing a strong preference for roads), provides little support for competitor avoidance or suppression within this landscape. One possibility is that ongoing baiting directed at foxes could be suppressing their numbers sufficiently to allow for temporal and spatial co-occurrence of cats.
For example, Johnson and VanDerWal (2009) demonstrated that the relationship between dingoes and foxes is likely to be triangular in shape (i.e. dingoes and foxes can co-occur, but dingo abundance generally sets the upper limit on the abundance that foxes can reach).
Another possible explanation is that in the relatively structurally complex landscape of our study, cats can easily retreat to shrubs or up trees if they encounter a fox.
A notable result of this study is that cats and (particularly) foxes appear to be widespread across the GNP despite an extensive and ongoing fox baiting programme. While this could suggest that the baiting programme is not achieving its intended aim of reducing the fox population to a sufficient level to alleviate predation on native wildlife, it is possible that occupancy is too coarse a metric for measuring success in management interventions, and it cannot reliably inform possible changes in population abundance. For example, lethal control may be reducing fox densities sufficiently to allow some predation relief on native prey species without significant reductions in site occupancy. Thus, while occupancy modelling can answer some questions, this approach is likely to overlook important relationships that require more detailed information (Nimmo et al., 2015). Other methods (e.g. spatial count, spatial-presence-absence), which can be used to measure predator densities in unmarked populations (i.e. where some or all individuals cannot be confidently identified), are likely to be more useful for teasing apart these types of relationships (Chandler & Royle, 2013;Ramsey, Caley, & Robley, 2015). These approaches have different survey requirements and assumptions (e.g. detectors must be spaced at a distance relative to the home-range of the target species so that a single individual is exposed to multiple detectors) and perform best with minimal bias when there are a high number of detections across each sample (i.e. >10) (Ramsey et al., 2015). So careful consideration must be given to survey design to ensure it is capable of answering the specific question at hand.
A primary challenge of monitoring programmes is ensuring there is adequate power for detecting effects of varying magnitudes (Guillera-Arroita & Lahoz-Monfort, 2012). This is further complicated by the fact that species vary in their detectability, distribution and abundance across landscapes. Improving power can be achieved by increasing the sampling effort (i.e. the number of sites or survey nights); however, this is typically limited by financial and logistical constraints (Field, Tyre, & Possingham, 2005;Joseph, Field, Wilcox, & Possingham, 2006) used to inform carnivore monitoring (but see Ramsey et al., 2017;Travaini et al., 2010;van Hespen et al., 2019).
Another crucial consideration is that as occupancy declines, our ability to detect a response also declines (as we have shown here), and so too might the detectability of the target species. When occupancy and detectability probabilities are low, more survey effort is required to detect a decline of a given magnitude, and so the choice of survey design becomes even more important.
While on-road sampling is likely to be an effective approach for determining whether carnivore populations are declining in response to management interventions, it too has its limitations. Little inference can be gained about the functional role of introduced carnivores across the entire landscape; for example, the presence of a cat or fox at a road site does not necessarily mean that they are using the adjacent vegetation. Restricting sampling entirely to roads and tracks may also limit the opportunity for concurrent predator-prey monitoring, and in turn, our ability to gain insight into predator-prey relationships. This is a fundamental question, given that predator control programmes are often implemented to protect native species. Answering these questions will likely require additional and complementary approaches (such as by combining diet, movement and camera trapping studies, see Hradsky, 2016), far greater survey effort and targeted monitoring across the entire landscape. Without some measure of impact (i.e. do introduced predators reduce native prey populations), or response (i.e. is the management intervention achieving its' intended aim of protecting native species), such lethal control programmes are difficult to justify (van Eeden, Dickman, Ritchie, & Newsome, 2017).
Nevertheless, our results suggest that surveys targeting roads -especially when resources may be limited -can be an efficient approach for determining if landscape-scale lethal control is effective.
We urge others to consider the importance of this for other ecosystems where carnivore monitoring and management occurs.

AUTHORS' CONTRIBUTIONS
HMG, EGR, MS, DGN, RD and BT conceived the ideas and designed the methodology; BT and DS collected the data; HMG analysed the data and led the writing of the manuscript. All authors contributed critically to the drafts and gave final approval for publication.

ACKNOWLEDGEMENTS
We would like to thank the traditional custodians of the Grampians-Gariwerd National Park, the Djab Wurrung and Jadawadjali people, for their ongoing support of scientific research on their land. This project was funded by Parks Victoria and the Victorian Government's Weeds and Pests on Public Land Program, which helps to ensure that Victoria's natural environment is healthy, valued and actively cared for. We thank Parks Victoria, and particularly Jessica Sharp for providing extensive logistical support. The preparation of this paper (including data analysis) was supported by the Australian Government's National Environmental Science Program through the Threatened Species Recovery Hub. We also thank three anonymous reviewers whose comments substantially improved this manuscript.

CONFLICTS OF INTEREST
The authors declare no conflicts of interest.

PEER REVIEW
The peer review history for this article is available at https://publons.