Using multiple data sources provides density estimates for endangered Florida panther


Correspondence author. E-mail:


  1. To assess recovery of endangered species, reliable information on the size and density of the target population is required. In practice, however, this information has proved hard to acquire, especially for large carnivores that exist at low densities, are cryptic and range widely. Many large carnivore species such as the endangered Florida panther Puma concolor coryi lack clear visual features for individual identification; thus, using standard approaches for estimating population size, such as camera-trapping and capture–recapture modelling, has so far not been possible.

  2. We developed a spatial capture–recapture model that requires only a portion of the individuals in the population to be identifiable, using data from two 9-month camera-trapping surveys conducted within the core range of panthers in southwestern Florida. Identity of three radio-collared individuals was known, and we incorporated their telemetry location data into the model to improve parameter estimates.

  3. The resulting density estimates of 1·51 (±0·81) and 1·46 (±0·76) Florida panthers per 100 km2 for each year are the first estimates for this endangered subspecies and are consistent with estimates for other puma subspecies.

  4. A simulation study showed that estimates of density may exhibit some positive bias but coverage of the true values by 95% credible intervals was nominal.

  5. Synthesis and applications. This approach provides a framework for monitoring the Florida panther – and other species without conspicuous markings – while fully accounting for imperfect detection and varying sampling effort, issues of fundamental importance in the monitoring of wildlife populations.


An accurate understanding of population status is fundamental for the management and recovery of endangered species (Campbell et al. 2002; Hoekstra et al. 2002). However, estimates of population size and density are lacking for many of the world's most endangered species. As a result, it has been difficult to quantify extinction risk and monitor the effects of conservation actions.

The Florida panther Puma concolor coryi is the last remaining puma subspecies in eastern North America. Originally occurring from Arkansas and Louisiana to South Carolina and Florida (Young & Goldman 1946), the current distribution is restricted to about 10 000 km2 in southern Florida (Kautz et al. 2006). Due to unregulated hunting in the 19th century and large-scale loss of habitat during the 20th century (Onorato et al. 2010), Florida panthers were listed as endangered in 1967 (US Federal Register 1967) and subsequently protected under the Endangered Species Act of 1973 (Public Law 93-205). Nevertheless, by the early 1990s, their population had dwindled to 20–30 individuals (McBride et al. 2008). Intensive population management, including introduction of wild-caught pumas from Texas to alleviate effects of inbreeding (Seal 1994; USFWS 1994), legal protection (O'Brien & Mayr 1991; Janis & Clark 2002), efforts to reduce road mortality (Foster & Humphrey 1995), and habitat and prey conservation (Janis & Clark 2002) have led to an increase in panther abundance (McBride et al. 2008) and genetic diversity (Johnson et al. 2010). Still, the Florida panther remains one of the most endangered felids world-wide (Onorato et al. 2010).

The Florida Fish and Wildlife Conservation Commission (FWC), with assistance from the federal government (e.g. National Park Service – NPS, U.S. Fish and Wildlife Service – USFWS), commenced research on the Florida panther in 1981, resulting in publications covering a variety of topics including: estimates of demographic parameters, habitat selection, assessment of genetic restoration and documentation of biomedical issues (Beier et al. 2003; Onorato et al. 2010). Despite the intensive research effort, producing rigorous estimates of population size for the Florida panther has eluded scientists for decades (Beier et al. 2003), yet abundance remains a central tenet of the USFWS recovery plan objectives (USFWS 2008).

Large, elusive carnivores such as pumas are typically difficult to sample, and accurate estimates of population-related parameters are often challenging to obtain. Obstacles include low sample sizes due to rarity, wide-ranging behaviour and concerns about invasive sampling methods. Mark–recapture techniques are generally considered the gold standard for generating robust estimates of population parameters. For many felid species, camera-trapping is increasingly used for abundance estimates because the technique is non-invasive and efficient. The resulting data, in combination with traditional capture–recapture (CR) models (e.g. Otis et al. 1978) or spatial capture–recapture (SCR) models (e.g. Efford 2004; Royle & Young 2008), have largely facilitated the estimation of demographic parameters of many felid species with unique pelage patterns (e.g. Karanth & Nichols 1998; Karanth et al. 2006; Royle et al. 2009). Although some puma studies use this combination of methods (Kelly et al. 2008; Negrões et al. 2010), the species generally lacks clear features for individual identification from photographs, seemingly rendering camera-trapping an unfeasible option for capture–recapture modelling of Florida panthers.

Alternatively, scat sampling in combination with genetic analyses can provide capture–recapture data (Royle, Kéry & Guélat 2011). Although this sampling technique has been applied in the study of felid populations (e.g. Ruell et al. 2009; Gopalaswamy et al. 2012), it would be difficult to implement for the Florida panther due to the subspecies' low genetic diversity (Roelke, Martenson & O'Brien 1993) and the fast decay of DNA in Florida's warm and moist climate (Farrell, Roman & Sunquist 2000; Lucchini et al. 2002).

Given the obstacle of individual identification, collecting capture–recapture data would require that animals be physically marked and recaptured. The high cost and safety issues to both animal and handler make such approaches impractical for elusive and potentially dangerous animals like large carnivores. This risk is compounded when dealing with the small populations of endangered species. Thus, non-invasive sampling techniques are preferable whenever possible (Long et al. 2008).

Florida panthers have been extensively studied using traditional very high frequency (VHF) and Global Positioning System (GPS) telemetry (e.g. Land et al. 2008; Onorato et al. 2011). Potentially, telemetry collars permit individual identification based on collar characteristics (e.g. different brands on different individuals or modifying collars with unique marks) observable in photographs. Under these circumstances, camera-trap surveys concurrent with existing telemetry studies can provide data suitable for population estimation in the framework of mark–resight models (e.g. Rice & Harder 1977; McClintock et al. 2009; McClintock & White 2010), which do not require that individuals be physically captured multiple times. Rather, a sample of individuals is captured and marked during a single marking event that occurs prior to resighting surveys, and a non-invasive technique such as camera-trapping or visual resighting can be used to collect ‘recapture’ data on these individuals. While mark–resight models provide robust estimates of abundance, they suffer from the same shortcomings as traditional capture–recapture models when it comes to estimating population density. To estimate density, we need to define the area sampled. This generally relies on ad-hoc approaches, which renders density estimates somewhat arbitrary.

Our objective was to provide a rigorous and statistically sound density estimate for Florida panthers in the Picayune Strand Restoration Project area (PSRP). We used data collected during a 21-month camera-trapping study (Shindle & Kelly 2007) and telemetry data simultaneously collected by the FWC in a new modelling framework that, analogous to traditional mark–resight, allows for only a portion of the population to be identified (Chandler & Royle In Press; Sollmann et al. 2013). Further, analogous to SCR models, this new framework explicitly links abundance to a clearly defined area, thus providing unambiguous density estimates. To improve the estimation of model parameters associated with individual location and movement, and to produce more precise estimates of density, we extend the model by also incorporating telemetry location data. We confirm the reliability of model results using a simulation study. Providing a rigorous estimate of Florida panther density, this modelling approach has wide application for animal conservation and endangered species management.

Materials and methods

Study area

The study was conducted in PSRP, an area that encompasses the former Southern Golden Gate Estates subdivision development, covering approximately 241 km2 in Collier County, Florida. Originally slated for housing development, the area is currently undergoing vegetative and hydrological restoration (U.S. Army Corps of Engineers 2004). Together with two neighbouring reserves, the PSRP forms a large block of panther habitat in the core of the subspecies' range. The climate of the study area is that of a tropical savannah with distinct wet (May–October) and dry (November–April) seasons (Duever et al. 1985).

Camera-trapping and radiotelemetry

From 2005 to 2007, 98 camera traps (Digital CamTrakkerTM, CamTrak South Inc., Watkinsville, GA, USA) with passive infrared heat-in-motion detectors were deployed in PSRP for 21 consecutive months as part of a pre-restoration baseline survey for panther and white-tailed deer Odocoileus virginianus (Shindle & Kelly 2007). A grid with 2-km2 cells was overlaid on the study area, and one camera was placed within each grid cell (Fig. 1). Most cameras were deployed along roads or trails and secured to trees approximately 45 cm above ground. Cameras operated 24 h per day with a minimum 20-s delay between sequential photographs. Camera traps were checked every 21–28 days to retrieve images and ensure units were functioning.

Figure 1.

Picayune Strand Restoration Project area, Southern Florida, with camera-trapping grid used to survey Florida panthers between 2005 and 2007, and radiotelemetry locations for three collared panthers (stars, circles and triangles) used in the spatial mark–resight model as the marked portion of the population.

The FWC monitors Florida panthers in the PSRP and neighbouring areas using radiotelemetry. Locations were collected via aerial telemetry three times per week (see Land et al. (2008) for methods). Manufacturers of radiocollars included Telonics (Mesa, Arizona, USA), Advanced Telemetry Systems (Isanti, MN, USA) and Followit (Lindesberg, Sweden). Collars from different manufacturers have distinct physical features and therefore provided a visual means of individual identification of collared panthers from camera-trap pictures.

Mark–resight models require that all marking takes place before resighting. Here, we regard those panthers as the marked part of our population that wore radiocollars throughout one or both primary camera-trapping occasions (see below) and used the PSRP as part of their home range. Panthers that were collared during the course of a primary occasion were regarded as ‘unmarked’. Although some photographs of uncollared panthers could be attributed to individuals based on natural marks, many photographs of uncollared panthers were ambiguous. Since mark–resight models require that individuals can always correctly be identified as marked or unmarked, we treated all photographs of uncollared individuals as unmarked. For photographic records of uncollared individuals, we treated subsequent pictures at a given camera trap as independent if they were separated by at least an hour. Photographs that showed two (three, etc.) individuals were treated as two (three, etc.) independent records. We discarded pictures that we were unable to verify whether the individual was collared or not. We further excluded dependent kittens and juveniles from our analysis.

Data analysis

Spatial capture–recapture models

We analysed concurrent photographic and telemetry data, building on the SCR model for partially marked populations described by Chandler & Royle (In press). Generally, SCR combines a model for individual location and movement with a model describing detection by traps, using individual and site specific detection data (Borchers & Efford 2008; Royle & Young 2008; Gardner, Royle & Wegan 2009; Borchers 2012). In SCR models, we assume that each individual i has an activity centre, si, and that all si are distributed uniformly across the state space S, an area including the trapping grid, chosen large enough to include all animals potentially exposed to sampling. We assume that the number of records of individual i at trap j and occasion k, yijk, is a Poisson random variable with mean encounter rate λij, which is a decreasing function of the distance, dij, from trap j to the individual's activity centre si. Under a half-normal encounter rate model,

display math

λ0 is the baseline trap encounter rate at dij = 0 and σ is the scale parameter of the half-normal function.

To estimate N, the number of activity centres in S, we employ data augmentation (Royle, Dorazio & Link 2007). Let n be the number of observed individuals. Then this approach is equivalent to augmenting the observed data set with Mn ‘all-zero’ encounter histories or ‘hypothetical individuals’ that were never observed. N is estimated as the sum of an individual auxiliary variable, zi,

display math

where = 1,2,3…M and zi = 1 if the animal is part of the population and 0 otherwise. The prior probability of Ψ is uniform (0,1), which corresponds to a discrete uniform (0,M) prior probability for N. M is an arbitrary value set sufficiently large as to not truncate estimates of N. Density, D, can be derived by dividing N by the area of S.

Extension of the SCR model to a mark–resight situation

Chandler & Royle (In press) extended this model to a mark–resight situation, where only part of the population can be individually identified. Under these circumstances, the individual encounter histories yijk are partially latent – only yijk for the m marked animals are observed. For the unmarked individuals, we observe only the accumulated counts njk = ∑yujk, where = {+ 1,…, N} is an index vector of the NU unmarked individuals. Unobserved encounter histories are essentially missing data. Adopting a Bayesian framework and using Metropolis-within-Gibbs (MwG) Markov chain Monte Carlo (MCMC) sampling, we can update missing data using their full conditional distribution (Gelman et al. 2004, Ch. 11). For the yijk from unmarked animals, the full conditional is multinomial with sample size njk:

display math

The remaining model parameters are then updated conditional on the full set of encounter histories.

When the number of marked individuals, m, is known, estimating N reduces to estimating the number of unmarked individuals U. In this situation, M= size of the hypothetical unmarked population in S. By updating the latent encounter histories (see above), we assign records of unmarked individuals to some of these hypothetical individuals, so that their encounter histories are no longer ‘all-zero’.

In non-spatial mark–resight models, an important model assumption is that marked individuals represent a random subset of the population. This assumption is still required in spatial mark–resight, but additionally, the marked individuals must represent a random sample of individuals in the state space S. Here, we have only a small set of marked individuals (see results), and the telemetry information for these individuals indicates that they are distributed throughout most of S (Fig. 1).

Incorporating telemetry location data

We can relate the parameters of the half-normal encounter rate model to those of a bivariate normal movement model (Calhoun & Casby 1958), with mean = si, and variance–covariance matrix , where the variance in both dimensions is σ2 and covariance is 0. Under this model, σ can be related to a measure of how far individuals move (Reppucci, Gardner & Lucherini 2011). Ordinarily, these parameters are estimated only from the trapping data. Telemetry data, however, provide more detailed information on individual location and movement. By assuming that the Ri locations of individual i, li, are a bivariate normal (Normal2) random variable:

display math

we can estimate σ, as well as si for the collared individuals, directly from telemetry location data using their full conditional distributions within the MwG sampler. Under this formulation, σ and si for the collared individuals are no longer conditional on the resighting data y, but only on l. For the unmarked individuals, si are estimated as in conventional SCR, conditional on the encounter histories. The full MwG MCMC sampler can be found in Appendix S1 (Supporting Information).

Model application to Florida panther data

To account for the lack of demographic population closure over 21 months of camera-trapping, we defined two primary occasions, from 1 July 2005 to 31 March 2006 and from 1 July 2006 to 31 March 2007. Within primary occasions, we grouped data by month and accounted for the number of days each camera trap was functional each month, tjk, using λij * tjk/30. We limited telemetry data used in our model to the same time periods. To define S, we used a 15-km buffer from the outermost coordinates of the trapping grid and removed parts of the resulting rectangle that comprised ocean or islands. This resulted in an area for S of 1719·13 km2.

We ran three chains of the MwG sampler with 200 000 iterations each, discarding 10 000 iterations as burn-in using the software R 2.13.0 (R Development Core Team 2011). To check for chain convergence, we calculated the Gelman-Rubin statistic R-hat (Gelman et al. 2004) using the R package coda (Plummer et al. 2006). Values below 1·1 indicate convergence; in our results, all model parameters had R-hat <1·1. We report the posterior mean (± standard deviation), mode, and 95% Bayesian credible intervals (95BCI) for all parameters.


During the two primary occasions, we accumulated 43 890 trap days and obtained 445 photographs of Florida panthers. We discarded 137 pictures that we were unable to determine whether they belonged to a radio-collared individual or not and one picture of a collared panther that traversed the study area but was not resident (see Discussion for further treatment of this topic). Of the remaining photographs, 17 were records of identifiable radio-collared individuals and 290 pictures showed uncollared panthers (Table 1).

Table 1. Collared Florida panthers present in the Picayune Strand Restoration Project area and used as marked individuals in the spatial mark–resight model, total number of photographs and number of photographic records of these collared individuals in the two 9-month primary camera-trapping occasions
OccasionNo. collared individualsTotal number of picturesNo. pictures of collared individuals
  1. a

    One individual from year 1 was present again in year 2.


Three individuals met our requirements of being collared throughout one or both primary sampling occasions, with two collared individuals being present in one primary occasion only, while one was present in both occasions. For each collared individual, we accumulated an average of 99·5 (SD 10·6) telemetry locations per primary occasion (Fig. 1).

The posterior mean for the movement parameter σ was 4·45 (±0·11) km. The baseline trap encounter rate λ0 had a posterior mean of 0·09 (±0·02) expected photographs per 30 days. The posterior mean for population density D was 1·63 (±0·50) individuals per 100 km2 in year 1 and 1·66 (± 0·56) individuals per 100 km2 in year 2; for both years, the posterior mode was slightly lower, at 1·51 and 1·46 individuals per 100 km2, respectively. Posterior summaries of parameter estimates are given in Table 2.

Table 2. Posterior summaries of parameter estimates from a spatial mark–resight model applied to Florida panther camera-trapping and telemetry data from the Picayune Strand Restoration Project area, Florida. Density is estimated for two 9-month primary occasions (t)
Parameter UnitMean (SE)Mode2·5%97·5%
σ km4·45 (0·11)4·464·244·68
λ 0 Pictures per 30 days0·09 (0·02)0·090·060·14
N (= 1)individuals in S27·98 (8·54)251447
N (= 2)individuals in S28·59 (9·67)251351
D (= 1) individuals per 100 km21·63 (0·50)1·510·812·73
D (= 2)individuals per 100 km21·66 (0·56)1·460·762·97

Simulation study

To investigate potential bias and precision of our estimators, we generated 100 data sets consisting of both camera detection and telemetry location data under the same conditions observed for the surveyed panthers (i.e. with parameters equal to the posterior means obtained in our analyses, and the trapping grid, sampling effort, number of known individuals and telemetry locations equivalent to values in the actual field study). Across 100 data sets, parameters were estimated with low accuracy (relative root mean squared error (RMSE) 26–39%); only the RMSE of σ was low, at 3%. For N, the posterior mode presented a less biased estimator (relative bias 11–13%) than the mean (27–29%). For λ0 and σ, relative bias of the mean was 4 and 0·3%, respectively. Coverage of the true values by 95% BCI was between 92% and 99% for all parameters (see Appendix S2, Supporting Information).


Large felids such as the Florida panther are notoriously difficult to monitor. Low population densities and elusive behaviour often result in sparse data, requiring intensive sampling over several years. Camera traps are an ideal tool for the study of large and wide-ranging species, but inference from camera-trap data for populations that cannot be individually identified is limited. Mark–resight methods have long been used as an alternative to traditional mark–recapture studies (e.g. Rice & Harder 1977; Minta & Mangel 1989), but only recently has the concept of mark–resight modelling been extended to SCR models (Chandler & Royle In press). This development has made it possible to address a major problem facing wildlife managers who are in need of reliable density estimates for rare and elusive species without conspicuous natural marks.

Florida panther density

The density estimates of approximately 1·5 individuals per 100 km2 summarize the current state of knowledge on Florida panthers in PSRP. Historically, there have been no reliable estimates of abundance or density for the Florida panther (Beier et al. 2003). Although the density estimate by Maehr, Land & Roof (1991) of one individual in 110 km2 was considered reasonable, it lacked confidence intervals and could not be applied elsewhere (Beier et al. 2003). Similarly, counts based on physical evidence (e.g. tracks, scats; McBride et al. 2008) do not account for varying sampling effort, possible double-counting of or failure to detect individuals, and they lack the potential for repeatability due to a reliance on expert observers for accurate interpretation of panther signs.

Our density estimates fall within reported densities of pumas in other parts of their geographical range. Generally, the lowest puma densities of ≤1 individual per 100 km2 are found in the northern part of the species' range (e.g. Hemker, Lindzey & Ackerman 1984; Laundré & Clark 2003). Except for areas heavily impacted by poaching and logging, Central and South America generally harbour higher puma densities, ranging from just over 1 to almost 7 individuals per 100 km2 (Kelly et al. 2008; Paviolo et al. 2009; Negrões et al. 2010; Soria-Diaz et al. 2010). Given the tropical climate and habitat of Florida, and the fact that PSRP is still recovering from heavy anthropogenic impacts, our density estimates of approximately 1·5 panthers per 100 km2 are consistent with previous findings.

The panther population of PSRP most likely declined because of the severe habitat degradation caused by water management practices and direct human disturbance. However, PSRP has two neighbouring reserves, the Florida Panther National Wildlife Refuge (FPNWR) and the Fakahatchee Strand Preserve State Park, both of which have been protected for several decades. Compared with these reserves, PSRP probably has less suitable habitat. Indeed, until recently, the PSRP area was mainly used by dispersing male Florida panthers, and reproductive events in the area were rare (Shindle & Kelly 2007). Applying the bivariate normal model to telemetry data from VHF and GPS collared individuals in the neighbouring FPNWR showed that individuals at this site have smaller home ranges (average σ was 3·44 km based on seven individuals), which in carnivore populations is often linked to a higher population density (e.g. Dahle & Swenson 2003; Benson, Chamberlain & Leopold 2006). Most likely, individuals from neighbouring reserves are immigrating into the PSRP area as it recovers from the severe anthropogenic impacts and as panther populations in the neighbouring areas expand.

Reliability of estimates

The precision of density estimates from spatial mark–resight models depends on the number of marked individuals (Chandler & Royle In press). In the present study, photographic data on the small number of radio-collared individuals were particularly sparse (17 pictures total), but incorporating telemetry information about individual locations and movements increased the precision of our density estimate. According to our simulation study, although we can expect some positive small-sample bias in estimates of N, we also expect the true value to fall within the 95BCI. As a result, our modelling framework represents a promising tool for population monitoring of far-ranging, elusive species. For species that are studied extensively using radiotelemetry (Land et al. 2008; Onorato et al. 2011), the combination of traditional sampling techniques such as radiotelemetry with the increasingly popular methods of camera traps and SCR modelling (Royle et al. 2009) is likely to replace more traditional inference methods (Nichols, O'Connell & Karanth 2011). This approach is not limited to Florida panthers, but applies to other species that are not ‘naturally marked’ but can be tagged or otherwise recognized, and can also be applied to other types of spatial resighting data, such as point counts for birds or amphibians. With adequate sample size, telemetry locations are not necessary to estimate population size, so tags can be anything that permits identification.

Current spatial mark–resight models assume that marked individuals are a random sample from the total population of S. This means, ideally, defining S should be part of the study design and marking efforts should be spread evenly within S. In practice, that may often not be realistic. When marked individuals are not a random sample of S, but were taken from a smaller area, density estimates are likely negatively biased. Relaxing this assumption is the focus of current SMR model development.

Implications for future Florida panther research

Despite the progress made towards recovery in over 30 years of research, the Florida panther population continues to require close monitoring. Our method is an improvement over monitoring methods historically implemented for three main reasons:

  1. Our model enables researchers to use camera traps, which allow for non-invasive monitoring of Florida panthers in regions where they are also monitored by telemetry.
  2. The spatial mark–resight model provides a standardized analytical framework that accounts for imperfect individual detection and varying sampling effort, so that estimates of density across time and space are comparable.
  3. Our modelling approach provides estimates of uncertainty about density estimates. As such, we can fully assess whether a sampling design is yielding appropriate data to monitor the Florida panther population or whether sampling has to be modified (in terms of sampling technique, design and effort).

Still, there is room for improvement. A basic assumption of any mark–resight approach is that the marked individuals are a representative sample of the population (McClintock & White 2010). This is generally accomplished by applying a technique that is different from the resighting method to mark a random sample of individuals (Bowden & Kufeld 1995). While the methods for marking and resighting were distinct in the present study, the extremely low number of collared individuals may not be representative of the entire population. Considering the difficulties, risks and costs associated with capturing large felids, tagging a larger sample of panthers may be challenging. But even adequate coordination of marking and resighting would be an improvement. In the present study, marking and resighting occurred concurrently and individuals tagged within the primary camera-trapping occasions had to be treated as ‘uncollared’. By tagging animals ahead of the resight surveys, this loss of valuable data could be avoided.

Owing to the low number of collared individuals, we were unable to incorporate sex- or year-specific differences in movement and detection into our model. Differences in these parameters between males and females are known to be pronounced for large carnivores (e.g. Gardner et al. 2010; Sollmann et al. 2011). For Florida panthers, males are known to have larger home ranges than females (Onorato et al. 2010). Further, collared individuals were photographed more frequently during the second primary occasion, which could indicate higher trap encounter rates. Ideally, future studies should aim at collecting enough data to allow for the modelling of these effects.

The sparseness of the data also precluded any formal treatment of transiency. Transiency is a common issue in open population capture–recapture studies (e.g. Pradel et al. 1997). In closed population studies, formally, the presence of transient individuals violates the fundamental assumption of population closure and is therefore generally not explicitly addressed but ‘assumed away’. Only because we had radiotelemetry locations, we were able to identify one of the collared panthers in our study as a transient and we decided to remove that individual from the data set. We cannot apply such a correction to the uncollared individuals. By removing transients from the collared individuals but not the uncollared, the former are arguably no longer a representative sample of the latter, which may introduce some positive bias into the estimates of density. We found, however, that retaining the transient individual resulted in unreasonable estimates of the movement parameter σ (data not shown). Given the transient's large movements this is not surprising: when applying the bivariate normal movement model to individual sets of telemetry locations, σ for the transient was 3·5 times larger than for the remaining individuals. Within the spatial mark–resight model, the estimate of σ almost doubled when retaining the transient. While it is disconcerting that a single individual impacted estimates to such a degree, this is a consequence of the small data set, where one outlier has disproportionate effects on model outcomes. With an adequate sample size (i.e. larger number of marked individuals), presence of a single outlier would have a much smaller impact. Further, the problem could be avoided or diminished by shortening the sampling time frame to better approximate a closed population. Even if a transient is present, over a short time interval, its movements are unlikely to be so pronouncedly different from resident individuals, thus diminishing its effect on parameter estimates. Alternatively, with adequate sample size, or as information on the proportion of transients in the population accumulates over time, transiency could be addressed explicitly within the model, for example, using an individual covariate describing transiency state. Regardless of the approach, future study design for Florida panther population monitoring has to both strive for larger sample sizes and consider the assumption of population closure.

Finally, identifiability of individuals on pictures could be improved, for example, by increasing camera trigger speed to allow more centred subjects and by taking multiple pictures per camera-trapping event. We discarded 137 pictures from analysis because we were unable to tell whether an animal was wearing a collar or not. If individuals can at least be identified as ‘marked’ (but not to individual level), their data can still be included in mark–resight models (e.g. McClintock et al. 2009; Sollmann et al. 2013).

In spite of these caveats, spatial mark–resight models allow for the development of a standardized protocol that can be applied by different investigators and at different study sites without compromising the comparability of results. As such, these models provide a valuable population monitoring tool for wildlife species that are not consistently identifiable to the individual level. For Florida panthers, spatial mark–resight models could be the cornerstone of a distribution-wide survey protocol to estimate the density or size of the Florida panther population. This is a current research priority and will be indispensable in helping quantify the level of success conservation, and management measures are having at achieving recovery objectives outlined by the USFWS.


We thank the U. S. Fish and Wildlife Service, U. S. Army Corps of Engineers, and U. S. Geological Survey for financial support and the North Carolina Cooperative Fish and Wildlife Research Unit for assistance with funding administration. We thank two anonymous reviewers and the editor Paul Lukacs for helpful comments on earlier drafts of this manuscript. In addition, we thank the citizens of Florida that support panther conservation via the purchase of ‘Protect the Panther’ licence plates. The FWC data used in this project were collected using revenue accrued from the purchase of these plates.