Turtles are among the most endangered vertebrate groups on Earth with more than 60% of species already extinct or threatened with extinction (Lovich et al. 2018; Rhodin et al. 2018; Cox et al. 2022). Habitat fragmentation, loss, and degradation—working separately and synergistically with climate change—and overcollection are widely acknowledged as the major drivers of decline worldwide (Stanford et al. 2020). These concerns have led to increased focus and attention on population monitoring in management plans as an essential component of many conservation strategies. Monitoring is particularly important because delayed sexual maturity and low fecundity characterizing turtle life history make population recovery slow (Gibbons 1987; Enneson and Litzgus 2008; Willey et al. 2021). Therefore, long-term monitoring studies and population modeling are invaluable tools for recognizing trends and identifying effective intervention strategies (Congdon et al. 1993; Lovich et al. 2018; Howell et al. 2019).

In North America, the wood turtle (Glyptemys insculpta) is listed as endangered on the International Union for Conservation of Nature (IUCN) Red List (van Dijk and Harding 2011), is currently listed under the federal Canadian Species at Risk Act (COSEWIC 2018), has been proposed for listing under the United States Endangered Species Act (Center for Biological Diversity 2012), and is currently listed as special concern, vulnerable, threatened, or endangered in most states and provinces in which it occurs (Jones and Willey 2015). In Maine, it is listed as a species of special concern and is a priority 1 species of greatest conservation need in the latest Maine State Wildlife Action Plan (Maine Department of Inland Fisheries and Wildlife 2015). Recent models incorporating empirical data suggest that there is a slight to moderate decline for this species range-wide (Moore et al. 2022).

Maine is currently considered a stronghold for the species due to suitable upland habitats, relatively unimpaired stream habitats (Willey et al. 2022), and abundant nesting areas (Jones et al. 2018). It also represents the northern extent of the species' range within the United States. In their comprehensive northeastern status review, Jones and Willey (2015) make the point that Maine harbors “regionally significant populations” and—along with West Virginia and New Hampshire—has the largest proportion of nonimpaired habitats but conversely was the least surveyed area, making it a priority region for standardized surveys. Furthermore, the northeastern part of the species’ US distribution has garnered particular interest in more recent conservation management approaches because climate change projections predict that range shifts for this species will be northerly and efforts may focus or rely on these populations (Mothes et al. 2020). Mothes et al. (2020) identified Maine—along with Vermont, New Hampshire, and New York—as being where the majority of climate refugia exist for this species within the United States. Thus, population monitoring and other demographic research in these northeastern areas are likely to become even more central to future conservation plans.

When creating or modifying species conservation plans, it is essential to have recent and local status assessments (IUCN–Species Survival Commission [SSC] Species Conservation Planning Sub-Committee 2017; Willey et al. 2021). To date, relatively little published research has been done on wood turtles in Maine, with notable exceptions being Compton et al. (2002) and Jones and Willey (2013). Current population-level pressures—such as overcollection, nest predation, nest failure due to temperature and limited degree-days at higher latitudes (Compton 1999), anthropogenic influences on habitat, and stochastic weather events—all highlight the importance of regular assessments, as any or all of these can dramatically affect turtle population status over even a short time span (Willey et al. 2021). Moreover, these pressures may be highly regionalized in their presence, extent, and manifestation (IUCN–SSC Species Conservation Planning SubCommittee 2017; Stanford et al. 2020). For any given location, a current and accurate population status assessment may be used to generate a population viability analysis (PVA), which is a modeling technique that explores the risks and dynamics of small populations to assess their long-term stability. These PVAs may then be incorporated into conservation plans to design effective management practices for improving or maintaining existing population trajectories in the face of regionally shifting ecological challenges (Akçakaya and Sjögren-Gulve 2000). It should be noted that PVAs, like all modeling exercises, are subject to assumptions and flawed parameter inputs (see Gross 2005 for one possible example) and should be used in conjunction with other sources of information when designing management plans.

We present modeling results using data gathered from a longitudinal monitoring project in central Maine to conduct a population demography study and subsequent PVA. Our goal is to present a robust and comprehensive analysis of an understudied yet purported stronghold area that may someday serve as a climate refuge. These models may contribute to 1) future meta-analytical assessments that rely heavily on numerous smaller studies to create comprehensive species overviews for regional-level planning, 2) filling gaps in information necessary for state-level management plans, and 3) directing future research and conservation efforts in this area.

METHODS

Field Sampling. — This study uses a 5-yr capture–mark–recapture data set on a single population in an area that has been substantially modified by human activity and much of which is actively managed lands consisting of agriculture, mining, and recreation. Owing to the problem of overcollection, we are omitting detailed site information to protect this population. We initiated the study in 2016, a pilot year in which only 2 individuals were captured. From 2017 to 2021, we conducted regular surveys of the study site (approximately every 1–3 wks) during the spring and fall (April–May and mid-September–October). We also regularly collected data on radio-tagged turtles during the summer months (June–mid-September) less systematically than during surveys but with the occasional incidental location of nontelemetered turtles in both upland and stream areas. The site consisted of an 8-km stretch of river, with 5 1-km transects spaced throughout where our survey efforts were focused. Survey teams consisted of 2–4 people walking slowly through or adjacent to each stream transect. Captures outside of regular stream surveys, such as during the summer months when most turtles are in upland foraging areas, were also included in the analysis.

On first capture, all individuals were marked with a unique identification code by filing V-shaped notches into 3 or more marginal scutes (modified from Cagle 1939). We also recorded sex, shell dimensions (midline straight-line carapace length [CL] and width, midline straight-line plastron length and width, and maximum shell height), mass, and any notable injuries or malformations.

For the purposes of our analyses, turtles with a CL > 180 mm were considered sexually mature and, therefore, adults (Ross et al. 1991; Brooks et al. 1992; Daigle 1997; Tuttle and Carroll 1997; Saumure and Bider 1998). A subset of approximately 10–20 animals (the exact number varied among years) were part of a habitat selection study and received radio transmitters, which they had for variable amounts of time (i.e., less than 1 season if the transmitter fell off, failed, or extended out of range to 5 seasons with sequential battery replacement); we limited all location events for these animals to a single data point per year in our models. Similarly, nontelemetered turtles located more than once per season were also considered a single location event per year. In all years of study, we encountered only 4 hatchlings under varied circumstances, making interpretation of these sparse data difficult; therefore, they were excluded from our analyses.

Population Modeling. — We estimated apparent survival rates, population sizes, and population growth rates (λ) using Program MARK (White and Burnham 1999). We used a POPAN formulation of the Jolly-Seber model to estimate annual survival rates and population sizes and a Pradel-λ formulation of the Jolly-Seber model to estimate population growth rates (Jolly 1965; Seber 1965). We binned capture histories from each year into a single binary value of 0 or 1 to form encounter histories. For example, an individual captured in 2017, 2018, and 2021 received an encounter history of 11001. We ran models on a data set that included all individuals in our study to get population-wide estimates of demographic rates. However, because demographic rates are known to vary between both sexes and age classes of chelonians, we ran the same suite of models on data sets where the adults and juveniles were separated. We estimated the number of males and females within each population separately using MARK to prevent differences in capture rate from biasing population estimates (McKnight and Ligon 2017). To retrieve a continuous output of survival based on size, we used CL as a continuous covariate in a Cormack-Jolly-Seber model. Due to overall low sample sizes, it was necessary in these continuous covariate models to assume constant survival and recapture probabilities to reach model convergence. In order to compare the survival of our telemetered turtles against the full population, we conducted a Known-Fate-Analysis in Program MARK to estimate survival of these individuals.

For the POPAN formulations on the adult turtles, we considered a candidate set of 16 models. For both survival and recapture probability, we considered models with constant, time-varying, sex-varying, and time- and sex-varying rates. To prevent lack of model convergence due to overparameterization, we assumed constant emigration and immigration rates and population size (N) across years. For the Known-Fate-Analysis of our telemetered turtles, we considered a candidate set of 4 models (due to a lower number of parameters) using a binned annual encounter history. We then used the corrected Akaike information criterion (AICc; Akaike 1973) to rank the models and used model averaging to produce parameter estimates from the models whose AICc weight was greater than 0.01. We removed models that failed to converge from the model-averaging process. For the Pradel-λ formulation, we considered the same set of 16 candidate models. We used Program RELEASE within Program MARK to test the goodness of fit of our full data set to the assumptions of the Jolly-Seber models. These goodness-of-fit tests showed that while there were no violations across any of the first 2 tests, there was a significant violation within test 3 for the final transition year of the study (p = 0.026, 2020–2021). However, when running the goodness-of-fit tests on just the adult data, there were no longer any significant violations of the tests.

Population Viability Analysis. — We modeled population viability in the program VORTEX V.10.1 (Lacy and Pollak 2022). To parameterize the model, we included all parameter outputs from Program MARK that were modeled as part of this study. For any parameter estimates that we were unable to obtain from our own data, we instead used the estimates aggregated from the published wood turtle literature in Moore et al. (2022).

Baseline Model. — Beissinger and Westphal (1998) suggested that initial PVAs should be conducted with simple models and with relative extinction probabilities. Our baseline model attempted to use the fewest assumptions, a time period equal to approximately 4–5 generations for wood turtles, and an extinction threshold of a single individual remaining in the population. We assumed no density-dependent growth because evidence of density-dependent growth is rare in chelonians (Burke et al. 1998; Shoemaker et al. 2013).

Approximations of nest survival in our population were based on annual nesting surveys made between 2018 and 2021. Likely locations (e.g., natural beaches and natural and artificial depositions of sand and gravel) were surveyed and selected each year in mid- to late May just prior to the start of nesting season (Bougie et al. 2020; Willey et al. 2021). Up to 3 trail cameras were placed at each of the 8–16 potential sites to assist nest monitoring efforts on 3–5 of the transects. Nests were later identified to species using a combination of camera trap images and direct observation of shell fragments, intact eggs, or the presence of hatchlings.

Estimates of annual survival of hatchling and juvenile age classes were taken from the reported average population for a typical year produced by Moore et al. (2022). Estimates of apparent annual subadult and adult survival were taken from Program MARK. We produced a specified age distribution for our population by counting scute annuli on individuals. For any individuals where the scutes were too old or unreadable, we used the shell wear index method described in Jones (2009). To produce age estimates of individuals that we have not yet collected, we used a random draw with a normal distribution to pull values from a randomly generated data set that had the same mean and distribution as our data. We did not include any inbreeding depression because evidence of inbreeding in chelonians is uncommon (Kuo and Janzen 2004; Mockford et al. 2005; Pittman et al. 2011). Table 1 contains a full list of the parameters and their sources used in the baseline model.

Table 1. Parameters used in a baseline population viability analysis of a Maine population of wood turtles along with the sources of those estimates.

View Fullscreen

Sensitivity Analysis. — We performed a sensitivity analysis to illustrate the impacts of slight perturbations to selected demographic rates on population persistence. We analyzed all demographic rates that we took from Moore et al. (2022) and those that we hypothesized would have the greatest impact on population viability if conservation strategies (e.g., caging of nests and head-starting) were used to manage this population. In addition, we also produced a model where we increased survival across age classes to create target survival estimates that would result in a stable population. For our sensitivity analysis values, we retrieved demographic rates from Moore et al. (2022) using the expert-averaged values for good, average, and bad years.

RESULTS

Our final data set consisted of 68 unique individuals captured from 2016 to 2021 (Table 2). In our 2016 pilot season, prior to formal surveys, we captured 2 individuals. In all years combined, we newly captured 15 males, 20 females, and 33 juveniles. The larger juvenile cohort included turtles on the verge of adulthood (e.g., at least 6 transitioned to adult size over the course of the study). New captures decreased over time but seemed to stabilize, while recaptures of nontelemetered turtles appeared similar across years (Table 2). We affixed radio transmitters to 47% of the turtles at the time of first capture, which they kept for variable amounts of time (some transmitters ultimately fell off or stopped working). We were able to positively identify 4–7 nests annually for a conservative total of 24 nests across the 4 yrs. All but 2 of these nests were depredated, yielding an 8% nest success rate.

Table 2. Captures of new live turtles and subsequent annual recaptures of turtles for each year of the study.

View Fullscreen

Population Modeling. — Across our entire population, time-constant apparent annual survival was 91.2% (95% CI = 83.9%–95.4%; Fig. 1) with an estimated total population size of 73 (95% CI = 69–85) individuals. However, annual apparent survival did vary across years in the most supported models, ranging from a low of 80.3% in 2019 to a high of 97.5% in 2018 (Fig. 1). The most supported Pradel-λ formulation included a time-constant adult and subadult population growth rate with λ of 1.14 (95% CI = 1.05–1.24). Using a time-varying λ model to acquire time-specific λ values shows that while λ fluctuates annually, there was no significant difference in estimated annual λ values across years, suggesting that the adult population has been in a continuous, slightly positive growth rate throughout the duration of our study. The Known-Fate-Analysis of our telemetered turtles estimated annual survival at 95.2% (95% CI = 73.0%–99.1%; Fig. 1). Our model including CL as a continuous covariate outperformed models without it and shows that as individuals reach larger body sizes, their apparent survival increases until reaching an eventual asymptote at around 180 mm when survival starts to level off in the upper-80thpercentile region (Fig. 2). While this model outperformed models that did not include CL as a covariate, due to small sample size and a subsequent lack of convergence in the model, we were unable to use this model type with more parameters to estimate λ values or population size.

View Figure

• Download as Powerpoint
• Download Figure

View Figure

• Download as Powerpoint
• Download Figure

When analyzing just the adult data set, the most supported model included a sex-varying apparent survival and recapture probability. The model-averaging procedure estimated that females had a slightly higher apparent survival of 94.9% (95% CI = 81.6%–97.8%) compared with males, which had an apparent survival of 92.8% (95% CI = 77.6%–97.4%; Fig. 1). The adult population size was estimated to be 39 (95% CI = 35–43) with a 92.4% (95% CI = 84.2%–96.6%) chance of annual encounter and a 14.5% (95% CI = 10.3%–19.6%) probability of entry into the population. In contrast, when analyzing the subadult data set, apparent survival of subadults was lower at 82.7% (95% CI = 68.6%–91.3%; Fig. 1). The subadult population size was estimated to be 38 individuals (95% CI = 33–43) with a 73.8% (95% CI = 55.2%–86.5%) chance of annual encounter and a 15.8% (95% CI = 11.4%–21.3%) probability of entry into the population. Again, due to small sample sizes, we were unable to reach numerical convergence to provide estimates of λ values for each sex or for juveniles.

Population Viability and Sensitivity Analyses. — The baseline PVA model estimated a λ of 0.93 (95% CI = 0.91–0.95) and a probability of extinction within 150 yrs of 100%. Despite some broad perturbations in our baseline demographic parameters (e.g., increase in nest survival from 8% to 84.3%), there were no scenarios included within our sensitivity analysis that increased the population's instantaneous rate of increase to greater than 1 when just a single demographic parameter was varied (Table 3). To reach a stable population trajectory (r = 0.001), survival needed to be increased across multiple demographic classes (Table 3; Fig. 3). For example, increasing apparent adult female survival by just 3.9% led to a roughly 20% increase in the population's instantaneous rate of increase, r, from –0.083 to –0.066.

Table 3. Parameters used in the sensitivity analysis of a Maine wood turtle population viability analysis.

View Fullscreen

View Figure

• Download as Powerpoint
• Download Figure

DISCUSSION

The demographic values that we estimated for this Maine population suggest there is high adult survival that is slightly higher but not significantly different from the average literature value of 90.8% (Moore et al. 2022). Other studies have found that to achieve population stability, adult survival needs to be > 95% (Compton 1999), which also falls within our confidence limits. Similarly, our estimate for subadult survival (82.7%) was very close to the average survival of juveniles aged 6–12 yrs (82.4%) estimated by Moore et al. (2022). In contrast, our observed rate of nest survival was only 8%, whereas data summarized by Moore et al. (2022) placed an average year of nest survival at 31%, or nearly 4 times our observed average. Even in a bad year, as classified by Moore et al. (2022), nest survival was estimated to be 22%, or nearly 3 times our observed average. While a substantial portion of our captures were juveniles, as defined as those with a straight-line CL of < 180 mm (49%; 33 of 68), many of these were older juveniles and subadults. We did not capture many individuals from the youngest age classes, which is suggestive of low recruitment and is a problem encountered in other studies as well (Willey et al. 2021). Therefore, we were unable to estimate survival in early life stages for modeling or comparison to other populations. While our full data set did significantly violate test 3 (i.e., the assumption that all marked animals alive at i have the same probability of surviving to i + 1) within the goodness-of-fit tests, it is unlikely that this violation is biologically meaningful and is more likely attributable to survival differences between the adults and juveniles, as demonstrated by the lack of significant violations when testing goodness-of-fit on just the adult data set.

Our baseline PVA model predicted an average λ of 0.93 (95% CI = 0.91–0.95) for this population, which is not significantly different from the λ of 0.912 estimated by Moore et al. (2022; Table 3) for an average population of wood turtles in the middle of their range. An unavoidable conclusion of a λ value less than 1 is a declining population that, without future changes to demographic rates, will end in extinction. While our population did have high estimates of apparent annual adult survival (3.6%–4.8% higher than estimated for an average year in an average population), it was not high enough to counter the extremely high rates of nest mortality recorded at our study site each year. Similar to the findings of many other studies (e.g., Congdon et al. 1993, 1994; Compton 1999; Enneson and Litzgus 2008; Crawford et al. 2014; Howell and Seigel 2019), changes to adult apparent survival rates, especially apparent adult female survival rates, had the most impact on population growth rate relative to the percent perturbation to the parameter. The differences in the estimated λ values from Program MARK for our adults and subadults (1.14, 95% CI = 1.05–1.24), along with our baseline PVA value for the entire population, indicate that this population is likely slowly declining and that, despite a stable and slowly growing adult population, this will remain the case without sufficient recruitment and higher juvenile survival. This conclusion is supported through recent modeling work by Bougie et al. (2022). Those authors found that head-starting and nest protection alone did not lead to population stability and that improving adult survival was important as well. Their baseline population growth rate, however, was only 88%, compared with our ∼ 94%. We believe a multipronged management strategy aimed at multiple age classes is required to achieve stability in this population.

We modeled the population as an open population, allowing for dispersal into and out of the system. Therefore, emigration into the population from subadults and adults may be responsible for differences in λ values between the two models. It is likely that, while recruitment into this population from breeding is low, immigration of subadults and adults into the system from adjacent populations has occurred and is partially responsible for the positive λ values. Our estimated Pent (probability of entry) value for adults was 14.5%, suggesting that immigration into the population, while somewhat rare, does occur, although additional years of monitoring would improve confidence in this value.

Sensitivity analysis is a common tool used in conservation to assess the potential impact of management strategies on a targeted population by exploring the demographic effects of adjusting various vital rates (Heppell 1998; Manlik et al. 2018). Our sensitivity analysis demonstrated that when individually permutating demographic values to what Moore et al. (2022) recorded as an expert value for a good year, there was no permutation that produced a stable or increasing population. However, when all demographic values were changed to reflect the expert value for a good year, our population's r-value increased to 0.013 (SD = 0.162), and there was a 0% chance of extinction (Fig. 3). We also used this analysis to assess the impact of increasing nest survival from our baseline value of 8% to a high of 84.3% to represent the potential increase in nest survival by researchers finding and protecting nests. However, our sensitivity analysis indicated, in agreement with other studies examining the impact of management strategies on turtle populations (Heppell 1998), that increasing nest survival alone may not be enough to produce a stable or increasing population (r = –0.054). Because the adult and subadult apparent survival is already high in this population, any management strategy designed to produce a viable population will likely need to include both mitigation of the high rates of nest mortality (e.g., using predator exclusion devices) and increased juvenile survival (e.g., head-starting).

Although outputs from PVAs should be analyzed and used cautiously in wildlife management (Caughley 1994; Beissiner and Westphal 1998; Manlik et al. 2018), due to a general increase in compounding uncertainty that may lead to overly pessimistic scenarios (Brook et al. 2000; Herrick and Fox 2012; Rueda-Cediel et al. 2015, 2018), a suite of studies have found well-developed PVA models to be an effective predictor of long-term changes in population abundances (Brook et al. 2000; Manlik et al. 2018). We took multiple parameter values from Moore et al. (2022) that collated the published literature or estimated rates based on expert opinion. While these are currently the best estimates, we acknowledge that they may not accurately reflect those parameters in our population. PVAs are best used as a starting point for the implementation of management strategies rather than as a deterministic prediction for the population. Our results indicate that focusing on early life stages is important to improve both our model estimates and the trajectory of our population. It should be noted, however, that this management strategy should accompany those based on conventional wisdom that improving or maintaining high adult survival is essential to population stability (Mullin et al. 2020; Bougie et al. 2022). Our PVA demonstrated that this population is likely declining and, at its current trajectory, has a 100% chance of experiencing extirpation within the next 150 yrs due to a lack of recruitment into the population.

This study makes available an analysis of a population that is in an area identified as a potentially important climate refuge (Mothes et al. 2020). Furthermore, these data, when combined with other state-level wood turtle work, provide insight into whether Maine should continue to be considered a stronghold for this species. We suggest that this consideration be revisited, at least for parts of the state where wood turtles face threats similar to those encountered in other parts of their range where populations numbers have declined. These threats include heavily anthropogenically impacted areas resulting from a working landscape, such as agriculture and extraction practices. Currently, protection is limited (Jones and Willey 2015), and listing under the US Endangered Species Act a decade later remains under review. A more informed perspective on species trajectory could have implications for policies, prioritizing initiatives, funding, and efforts by both individual researchers and state and federal agencies. Jones and Willey (2015) also noted that decreasing numbers of wood turtles, leading to population collapse, are essentially irreversible without expensive and intensive management. Since turtle populations may take an exceptionally long time to recover from population declines due to long generation times and a potential lack of density-dependent compensatory increases in fecundity (Heppell 1998; Keevil et al. 2018), it is imperative that declining populations be identified and management strategies implemented as soon as possible. Studies such as this one with longitudinal monitoring and demographic analyses of a representative population can reveal issues that may be occurring locally or region-wide and can inform and enable more timely mitigation responses and preventive measures. These findings reinforce the need for targeting multiple life stages with multipronged strategies, such as employing nest predator exclusion devices, headstarting, repatriating confiscated animals, enacting anti-poaching measures, and changing land use practices.