Spatially extensive wildlife monitoring is typically costly, time consuming, and logistically challenging (Jones 2011). This can be especially true for difficult to capture species in relatively inaccessible environments. Consequently, it is common for researchers to focus on a small number of sites to make inferences about what factors drive population changes, and for managers to make decisions or evaluate actions using data collected at only a few sites (Lindenmayer et al. 1993; Beissinger and Westphal 1998; Ellner et al. 2002; Gurevitch et al. 2016). However, drawing robust conclusions about population drivers or management effects of wide-ranging species necessitates monitoring many populations over a large spatial scale. Otherwise, the application of monitoring results may reflect an understanding of 1 or a few uncharacteristic sites.

Using counts of unmarked animals, rather than more costly capture–recapture data, is a common means of scaling up the number of sites that can be monitored with available resources. Binomial N-mixture models (Royle 2004) in particular have become a widespread tool to estimate abundance because they use repeated counts to correct for imperfect detection, similar to capture–recapture abundance models. Collecting count data is appealing because it may require less effort, training, time, or resources than methods requiring animals to be in hand for individual identification. However, N-mixture models are less robust than capture-based models to assumption violations and, thus, are more prone to bias (Duarte et al. 2018; Knape et al. 2018; Link et al. 2018). For example, the assumptions of closure and constant detection are difficult to meet in many studies and violations of these assumptions can bias estimates.

Because of the bias documented in abundance estimates from N-mixture models, these models may not be suitable to estimate abundance in many scenarios. However, N-mixture models (and other methods that can account for variable detection: e.g., Poisson regression) may still provide a useful index of relative abundance (Barker et al. 2018). Thus, whether count data and N-mixture models are a functional solution for large-scale monitoring is likely context dependent, making it necessary to investigate under what, if any, circumstances these methods can be used to inform a specific monitoring objective for a given species.

The diamondback terrapin (Malaclemys terrapin) is an estuarine habitat specialist that is challenging to monitor in its salt marsh habitat, which can be difficult to access and traverse. It is considered a high priority species or Species of Greatest Conservation Need in wildlife action plans throughout its range, which extends from Cape Code Massachusetts to Corpus Christi, Texas (Ernst and Lovich 2009). In Georgia, where available estuarine habitat consists of expansive tidal salt marshes, the diamondback terrapin is considered a high priority species (Georgia Department of Natural Resources Wildlife Resources Division 2015). Terrapin populations are threatened by multiple factors including vehicular road mortality during nesting, high nest depredation by subsidized predators, and coastal development (Isdell et al. 2015; Chambers and Maerz 2019; Maerz et al. 2018). However, the greatest present-day threat to terrapin populations is commercial and recreational crab fisheries because terrapins are attracted to, enter, and drown in crab traps (Roosenburg 2004; Grosse et al. 2009, 2011; Isdell et al. 2015; Lovich et al. 2018; Chambers and Maerz 2019).

The Georgia Department of Natural Resources (GA DNR) is interested in developing a long-term monitoring program that it can use to identify and track relative levels of terrapin abundance among creeks throughout the state and trigger varying intensities of management actions. Such management actions may range from recreational and commercial fisher awareness efforts to the requirement of bycatch reduction devices or restricting crabbing within specific areas of marshes. By implementing targeted, status-dependent management actions, the GA DNR aims to minimize the need for more time intensive, cost intensive, or restrictive actions while stabilizing and recovering terrapin populations.

Following an intensive, state-wide 2-yr study of 29 tidal creeks to assess the effects of commercial crabbing pressure and road proximity on terrapin abundance (Grosse et al. 2011), terrapin monitoring methods in Georgia have consisted of annual capture–recapture surveys at 2–4 creeks (e.g., Crawford et al. 2018; Bradke et al. 2024; GA DNR Wildlife Resources Division, unpubl. data, 2018–2023) and systematic sampling of nesting females along a single causeway to Jekyll Island (Crawford et al. 2014a, 2014b, 2018). Capture–recapture within tidal creeks is time and labor intensive and, therefore, difficult to implement over many sites and long periods. For example, sampling to estimate abundance in the 29 creeks studied by Grosse et al. (2011) took 2 yrs and cost approximately $250,000 to complete.

Unmarked methods (i.e., head counts of terrapins surfacing in their aquatic habitat) that do not require capture and individual identification have also been considered for monitoring terrapins in Georgia. A study by Harden et al. (2009) showed a positive correlation between number of terrapins counted during a head count survey and the number subsequently captured when seining the creek, though there was a high amount of uncertainty in the relationship. Harden et al. (2009) recommended refinement of head count survey methods as a means to improve terrapin monitoring. For example, Harden et al. (2009) found that significantly more terrapins were observed during counts conducted at low tide. Lovich et al. (2018) subsequently used head counts at low tide and a single detection covariate, cloud cover, to show that terrapin abundance among tidal creeks within the Savannah Coastal Refuges Complex, Georgia, was negatively correlated with the number of commercial crab pots at the refuge unit scale. Though Harden et al. (2009) and Lovich et al. (2018) incorporated tide and cloud cover and Harden et al. (2009) included other detection related covariates, a limitation of these studies is that they did not incorporate repeated samples to address the issue of imperfect detection during counts. In both studies, the authors summed head counts from 2 passes within a creek. Recently, a study in Wellfleet Bay, Massachusetts, demonstrated that repeated terrapin head counts within a relatively “closed” sample modeled under a binomial N-mixture model framework was potentially an improved method for monitoring terrapin abundance (Levasseur et al. 2019). The Wellfleet Bay study renewed interest in applying a similar methodology to meet terrapin monitoring objectives in Georgia. However, the relatively open bay environment differs from the narrow, winding tidal creeks occupied by terrapins in Georgia. Thus, further investigation is needed to determine whether the more limited fields of view and lower accessibility of sites could impede the use of this method. Additionally, the use of repeated head counts has not been evaluated in the context of actual management objectives.

The objective of this study was to evaluate the ability of repeated terrapin counts and the N-mixture model to meet large-scale monitoring and future management objectives. Despite the propensity of N-mixture models to produce biased estimates, such a monitoring method may still be useful to GA DNR as an index of relative abundance if important sources of heterogeneity in detection probabilities can be accounted for using covariates. However, the usefulness of estimates from head count surveys to GA DNR also depends on minimizing classification error so that tidal creeks are sorted into reliable categories of abundance. Before the methodology is used, there is a need to evaluate the robustness of relative abundance estimates from head count surveys and to understand how the timing of sampling within the active season, location of sampling along a given creek, and amount of survey effort (e.g., number of repeated counts per visit) affect estimates. Understanding how these factors affect which category of relative abundance a creek may be assigned to will inform optimal monitoring design and the ability of GA DNR to conduct informative monitoring given resource constraints and the agency’s other management obligations.

METHODS

Data Collection. —

We used a modified version of the protocol described by Levasseur et al. (2019). During 4 April to 13 July 2021 and 21 April to 13 July 2022 we conducted repeated head counts of terrapins in 58 tidal creeks spanning the entire Georgia coast, including multiple sites within a subset of tidal creeks (119 total sites; Fig. 1). Each creek was surveyed 1 to 5 times per year (in at least 1 yr) and included 1 to 6 sites, resulting in 326 surveys of 112 sites in 2021 and 280 surveys of 83 sites in 2022. Thus, in total, we conducted 606 head count surveys (i.e., 606 unique site-visit combinations). To target times when terrapins would be present within creeks versus high marsh habitat, we conducted all surveys within approximately 3 hrs of low tide when the water line was below the high marsh (delineated by the presence of smooth cordgrass; Spartina alterniflora; Harden et al. 2009). Like Levasseur et al. (2019), we conducted each survey from a fixed survey point and used a rangefinder to determine the sample area. The area sampled extended up to 100 m from the observer in any direction; however, because creeks varied in width and sinuosity, which influenced visibility, the area sampled differed among sites (average: 6104 m²; range: 739–21,552 m²). During each survey, we performed 5–10 repeated scans of the sample area (mean = 9.93; SD = 0.53), with 1 min between each scan to allow for mixing by individuals (Levasseur et al. 2019), and we recorded the number of terrapin heads counted during each scan. Most surveys were conducted by a single observer, but a subset of surveys included 2 or 3 observers. In cases of more than 1 observer, each scan by each observer was treated as a replicate (i.e., counts were never pooled).

View Figure

• Download as Powerpoint
• Download Figure

When conducting each survey, we recorded variables that we hypothesized may explain terrapin detection or abundance, including the number of crab pots visible at each site, the average wind speed (m/sec; based on recording wind for approximately 1 min) using a kestrel meter, salinity (ppt) using a traceable salinity pocket tester (Cole-Parmer model EW-19601-21), cloud cover (no clouds, partial cloud cover < 50%, ∼ 50% cloud cover, partial cloud cover > 50%, overcast), and the perceived effect of waves and glare on the observer’s ability to observe terrapins (no issue, minor issue, major issue). We also used geographic information system (GIS) software and spatial layers to measure additional abundance covariates. These variables included the size of each area sampled (m²) and the distance of sample area to nearest potential nesting habitat. To identify potential nesting habitat, we used a digital elevation model produced using LiDARdata for the National Oceanic and Atmospheric Administration (NOAA; NOAA Office for Coastal Management 2012). Based on overlaying the digital elevation model on top of satellite imagery of the Georgia coast we established that the minimum elevation where terrapins are known to nest was 1.5 m. Therefore, we deemed any land surrounding the marsh that was ≥ 1.5-m elevation to be potential nesting habitat and measured straight-line distance from the edge of each sample area to this habitat. We also calculated the proportion of the shoreline that was armored (i.e., classified as bulkhead, revetment, or offset) within 100 m of sample areas. To do so, we considered the shoreline to be the perimeter of all land ≥ 1.5-m elevation, and we determined the proportion of that land that was armored using shoreline armoring spatial data developed by the Skidaway Institute of Oceanography from The Georgia Wetlands Restoration Access Portal (Alexander 2015).

We determined if crabbing was known to occur within or near each site using observations of crab pots recorded during our data collection and using all additional records of crabbing within Georgia that we were able to obtain. These records included seining surveys (2007–2008; Grosse et al. 2011), drone surveys (J. Maerz, unpubl. data, 2009), and crab pot surveys conducted by the Department of Natural Resources (GA DNR Wildlife Resources Division, unpubl. data, 2004–2009). We combined all datasets to determine whether a creek either had known crabbing activity or did not have known crabbing activity.

Model Selection. —

Using a Bayesian framework, we fit a binomial N-mixture model (Royle 2004) to repeated count data, with each scan of the sample area representing a replicate count, to estimate detection probability (p) and abundance (N). Like Levasseur et al. (2019), we used a static model where we modeled each of the 606 unique site-visit combinations as if it was an independent site. We modeled each count (y) at site-visit (i) during scan (k) as a binomial outcome conditional on abundance (N_i) with detection probability (p_i,k), and we modeled N_i as a Poisson outcome with mean equal to expected abundance (λ_i): $$N_i\sim \mathit{Poisson}({\lambda }_i)$$$$y_{i,k}|N_i\sim \mathit{Binomial}(N_i,p_{i,k})$$

Because our data likely violated N-mixture model assumptions (i.e., closure and constant detection) we considered N an estimate of relative abundance rather than absolute abundance.

To select variables important in explaining detection probability and relative abundance, we modeled the replicate count data from all surveys across both years and used indicator variable (IV) selection (Hooten and Hobbs 2015). We constrained λ_i using a log link and included the following covariates to explain this parameter: sample area, known crabbing activity, proportion of shoreline armoring, distance to potential nesting habitat, mean salinity (averaged by creek across all dates sampled), and day of year (including main effect and quadratic terms). Because diamondback terrapins are known to exhibit high site fidelity to specific tidal creeks, we included a random effect of creek on abundance to account for nonindependence among all sites and surveys within the same creek. However, we did not include a random effect of site to account for nonindependence among multiple surveys at the same site. Sample area and the random effect were included in all models (not varied during IV selection). We also included 4 covariates to explain detection probability, which we assessed using IV selection. These covariates were average wind speed, cloud cover, perceived effect of waves, and perceived effect of glare. We constrained p using a logit link and these 4 covariates.

To implement IV selection, we used “slab and spike” priors (Hooten and Hobbs 2015; Lawson et al. 2022) and multiplied each beta coefficient associated with covariate x (β_x) by a binary IV (w_x). Each IV was drawn from a Bernoulli distribution with probability p.w_x, which was given a uniform prior between 0 and 1. Beta coefficients were given normal priors with mean of 0 and variance equal to (1 – w_x) × (1 ÷ Spike) + w_x × (1 ÷ Slab), where Spike indicates the value used for spike precision (1000) and Slab indicates the value used for slab precision (0.01 for abundance covariates or 0.37 for detection covariates). The proportion of Markov chain Monte Carlo (MCMC) iterations where w_x = 1 indicated the amount of support for the covariate in the model. We derived model weights for all possible combinations of IV selection covariates (1024 models), where weights were equal to the proportion of MCMC iterations in which the covariate combination appeared. Because we were not interested in all possible combinations of covariates, we redistributed model weights to compare only models of interest (432 models). Models eliminated from model selection included all models with the quadratic term for day of year but not the associated main effect, because these models were not expected to be biologically meaningful. We also eliminated models that had both wind speed and wave effects or both cloud cover and glare effects because each of these pairs of variables were considered 2 different measurements of the same effect. We selected the top-ranked model (model with highest weight) for estimating relative abundance (described below).

We standardized covariates by centering on the mean and scaling by standard deviation prior to analysis and fit the model in JAGS (Plummer 2003) with the jagsUI package (Kellner 2021) in program R (4.2.1; R Core Team 2022). We used vague prior distributions on all parameters not included in IV selection (i.e., mean detection probability, mean expected abundance, standard deviation of random effect of creek on abundance, beta coefficient for sample area effect on abundance; Supplemental Table 1). We used 3 MCMC chains with 350,000 iterations each and burn-in of 50,000. We assessed chain convergence visually using trace plots and with Gelman and Rubin (1992) diagnostic values (i.e., Ȓ < 1.1).

Table 1. Relative abundance categories (< 1 terrapin per 6000 m², 1–10 terrapins per 6000 m², and/or > 10 terrapins per 6000 m²) occupied by each site among all visits within the same year (for sites visited ≥ 2 times). Sample sizes (n) indicate the number sites with the respective number of visits during each year of the study. Data were collected 4 April–13 July 2021 and 21 April–13 July 2022 within 119 total sites spanning the Georgia coast.

View Fullscreen

Estimation and Categorization of Relative Abundance. —

To obtain estimates of relative abundance across all sites and visits, we ran the top-ranked model identified by IV selection using vague prior distributions (Supplemental Table 1; all supplemental material is available at http://dx.doi.org/10.2744/CCB-1623.1.s1). We again used 3 MCMC chains with 350,000 iterations each and burn-in of 50,000 and assessed chain convergence visually using trace plots and using Ȓ values. We assessed goodness-of-fit by calculating chi-square discrepancy statistics, computing a Bayesian p-value, and estimating c-hat (Kéry and Royle 2016).

To improve comparability and interpretability of relative abundance estimates, we scaled each estimate by 6000 m², which was the approximate size of our average sample area. A creek that was approximately 30-m wide and surveyed 100 m in each direction lengthwise would have a sample area of approximately 6000 m² (Fig. 2). We categorized estimates of relative abundance using 3 thresholds that could be used by the DNR to determine the intensity of management actions to implement within each creek. The lowest threshold was < 1 terrapin per 6000 m², representing sites that were virtually without terrapins. The intermediate threshold was ≥ 1 but ≤ 10 terrapins per 6000 m². The highest threshold was > 10 terrapins per 6000 m², representing moderate to high relative abundance. To be placed in the intermediate or highest abundance categories, we required a 95% probability that relative abundance met or exceeded the respective thresholds. In other words, 95% of the posterior distribution had to be at or above the respective threshold.

View Figure

• Download as Powerpoint
• Download Figure

In addition to the 3 categories described above, we also compared our estimates of relative abundance to estimates intended to represent a natural density of terrapins in creeks not under any crabbing pressure. To do this, we used estimates from Grosse et al. (2011), who estimated density of diamondback terrapins per 1 km of creek within 29 tidal creeks that experienced a range of crabbing pressure. To scale densities estimated by Grosse et al. (2011) to our estimates, we measured mean creek width of the Grosse et al. (2011) study sites using GIS software. We then multiplied the 1 km creek length used in the Grosse et al. (2011) study by the measured mean creek width (20 m), to infer an average sample area of 20,000 m². Scaling the Grosse et al. (2011) mean density estimates (range 100–361 per 20,000 m² area of creek) to our mean sample area yielded 30–108 individuals per 6000 m². Thus, we quantified the number of surveys of our sites with 95% of the posterior distribution ≥ 30 terrapins per 6000 m².

Because 13 of the 58 creeks we sampled were seined by Grosse et al. (2011), we also compared abundance estimates within these creeks. Again, we rescaled densities estimated by Grosse et al. (2011) to facilitate comparability between the studies. We approximated areas associated with 1 km length of creek for each of the 13 seining sites using the above mentioned GIS measurements of mean creek width (i.e., 1 km length x mean creek width) and rescaled the Grosse et al. (2011) density estimates for each of these areas to 6000 m².

Evaluating Sensitivity of Estimates to Timing, Location, and Effort. — To assess the sensitivity of relative abundance estimates to the timing of sampling within the active season, we selected all sites visited ≥ 2 times within the same year, and we quantified the percentage of time that each site was placed in each category of abundance. To evaluate whether relative abundance estimates varied based on location of sampling along the creek, we selected all creeks with ≥ 2 sites and quantified the percentage of sites surveyed on the same date within a given creek that were placed in each category of abundance.

To assess the effect of survey effort on creek categorization, we used a random subset of the full dataset, which included a single survey of each site each year. This sampling scheme is more representative of a plausible number of surveys for DNR biologists to conduct, given their time and labor constraints. Then we used either 10, 5, or 3 scans per survey to estimate relative abundance by fitting each of these 3 datasets to the top-ranked N-mixture model. We compared the proportion of sites placed into each relative abundance category across the 3 levels of survey effort (10, 5, or 3 scans). We used the same MCMC settings and model checking methods as described for the full-dataset analysis.

RESULTS

We conducted a total of 606 surveys across all sites, 71 of which had 2 or 3 observers. We detected terrapins at 235 out of 606 (39%) surveys representing 38 out of 58 (66%) creeks. In total, we counted 2202 terrapins (mean = 0.32 terrapins/scan; range = 0–17).

Model Selection. —

The top-ranked model using IV selection included a random effect of creek and fixed effects of sample area, distance to potential nesting habitat, salinity, and day of year (main effect and quadratic terms) to explain abundance and fixed effects of perceived effects of waves and glare to explain detection (Fig. 3, Supplemental Table 2). Trace plots and Ȓ values for the IV selection model indicated successful convergence.

View Figure

• Download as Powerpoint
• Download Figure

Estimation and Categorization of Relative Abundance. —

Of all 606 surveys, 424 (70%) were in the < 1 terrapin per 6000 m² category, 163 (27%) were in the 1–10 terrapins per 6000 m² category, and 19 (3%) were in the > 10 terrapins per 6000 m² category (Supplemental Fig. 1). Only 1 survey of 1 site (surveyed twice in 2021 and once in 2022) had an estimated relative abundance ≥ 30 terrapins per 6000 m². All relative abundance estimates within creeks that were sampled by Grosse et al. (2011) were lower in our study. Mean estimates for these sites within our study ranged from 0 to 39 individuals per 6000 m² and rescaled estimates from Grosse et al. (2011) ranged from 0 to 567 individuals per 6000 m².

Variation of Estimates with Timing. —

We visited 91 sites ≥ 2 times in 2021 and 73 sites ≥ 2 times in 2022. Among the 164 total instances where we visited sites ≥ 2 times within the same year, 66% of sites were consistently placed into the same category each time they were surveyed during the same year (58% < 1, 7% 1–10, and 1% > 10; Table 1). Of the remaining sites, 29% were placed into both the < 1 and the 1–10 categories and 5% were placed into both the 1–10 and > 10 categories (Table 1). Only 1 site was placed in all 3 categories across surveys within the same year (i.e., 4 surveys in 2021, Fig. 4A).

Variation of Estimates with Location. —

Out of 171 total cases where we visited ≥ 2 sites within the same creek on the same date, 72% had all sites placed within the same category (53% < 1, 19% 1–10, and 1% > 10; Table 2). Of the remaining creeks, 22% had 1 or more sites placed into each of the < 1 and the 1–10 categories, 3% had 1 or more sites placed into each of the 1–10 and > 10 categories, and 1% had 1 or more sites placed into each of the < 1 and > 10 categories (Table 2). Four creeks had sites that spanned all 3 categories across surveys on the same date (e.g. Fig. 4B), with 1 of these creeks consisting of 4 sites and the other 3 creeks consisting of 5 sites.

View Figure

• Download as Powerpoint
• Download Figure

Table 2. Relative abundance categories (< 1 terrapin per 6000 m², 1–10 terrapins per 6000 m², and/or > 10 terrapins per 6000 m²) occupied by each tidal creek across all sites within the same creek sampled on the same date (for creeks with ≥ 2 sites). Sample sizes (n) indicate the number surveys for which the respective number of sites were sampled within the same creek. Data were collected 4 April–13 July 2021 and 21 April–13 July 2022 within 119 total sites spanning the Georgia coast.

View Fullscreen

Trace plots and Ȓ values indicated successful chain convergence for the model used to evaluate sensitivity of estimates to timing and location (i.e., the top-ranked IV model fit to the full dataset). A Bayesian p-value of 0.14, c-hat of 1.04, and the scatter plot of discrepancy values (Supplemental Fig. 2) indicated adequate fit of the model to the data.

Effect of Survey Effort on Categorization. —

As the number of scans of the site decreased, the ability to confidently place sites into the 1–10 and > 10 relative abundance categories also decreased (Table 3). No sites shifted to a higher abundance category as less survey effort was used, but 14 sites shifted 1 category lower when scans decreased from 10 to 5 and 11 sites shifted 4 category lower when scans decreased from 5 to 3 (Table 3, Supplemental Fig. 3). Trace plots and Ȓ values indicated successful chain convergence for all 3 reduced dataset models. Bayesian p-values for the 3 models ranged from 0.10 to 0.22 and c-hat ranged from 1.09 to 1.14. Scatter plots of discrepancy values (Supplemental Fig. 2) indicated adequate fit of all models to the data.

Table 3. Percentage of sites assigned to each relative abundance category (< 1 terrapin per 6000 m², 1–10 terrapins per 6000 m², or > 10 terrapins per 6000 m²) based on amount of sampling effort. Sample sizes (n) indicate the number sites sampled each year. Data were collected 4 April–13 July 2021 and 21 April–13 July 2022 within 119 total sites spanning the Georgia coast.

View Fullscreen

DISCUSSION

We demonstrate the potential of diamondback terrapin head count surveys conducted within tidal creeks to meet Georgia’s state-wide monitoring objectives. While this survey method demonstrated potential in Wellfleet Bay, Massachusetts (Levasseur et al. 2019), the tidal creek system of Georgia differs substantially from the bay environment. Wellfleet Bay is a relatively open bay with good, generally unobstructed observer visibility, whereas the tidal creeks in Georgia are narrow and winding, creating more limited fields of view. However, our results suggest that head count surveys may also have utility within salt marsh tidal creeks, which compose a large amount of the estuarine habitat used by terrapins (Roosenburg and Kennedy 2019). We were able to use this unmarked repeated count method to categorize sites into 3 levels of relative abundance, while accounting for uncertainty in estimates, suggesting its potential to inform the implementation of management actions by the GA DNR.

The potential of the N-mixture method is also bolstered by our identification of covariates that affect variation in detection, which suggests that using an N-mixture model may be an improvement over methods that do not account for imperfect and variable detection. Results of model selection indicated the importance of an observer’s perceived effects of waves and glare on detection probability, with waves producing a significant negative effect and glare producing a significant positive effect. The direction of the relationship between glare and detection was unexpected but reasonable because, while glare may be uncomfortable for the observer, the conditions created by glare can make terrapin silhouettes more visible against the water. Accounting for factors important in explaining detection probability can make this method a more reliable index of abundance than methods that do not account for such factors (Barker et al. 2018).

Although the ability to categorize sites by level of abundance while accounting for uncertainty in the estimates is promising, this method has limitations. Because sites lacked barriers to prevent migration and were small compared to the movement ability of terrapins (Harden et al. 2007; Lamont et al. 2021), sites did not appear to be geographically closed during surveys. Even if the number of repeated scans was limited to three, some terrapins may enter or exit the site during a survey. Therefore, this survey method violates the assumption of closure implicit in binomial N-mixture models (Royle 2004). Using multiple simultaneous observers could help decrease closure violations by reducing the survey duration, provided there are enough personnel available. Data collected with closure violations result in estimates that have high bias but may be interpretable as the number of individuals associated with the site during the entire survey, rather than absolute abundance (Kéry and Royle 2016). However, a second assumption of N-mixture models, difficult to meet in any wildlife monitoring study, was also likely violated; the “constant detection” assumption conditions that there is no unmodeled heterogeneity in detection probability (i.e., detection is constant or is fully explained by covariates; Martin et al. 2011; Kéry and Royle 2016; Barker et al. 2018; Duarte et al. 2018; Knape et al. 2018). Individual heterogeneity cannot be accounted for in N-mixture models because individuals are not uniquely identified, although heterogeneity among individuals is likely ubiquitous in wildlife monitoring (Kéry and Royle 2016; Veech et al. 2016). It is also likely that there is unaccounted for heterogeneity between replicate counts, surveys, and sites, despite our efforts to include the important, measurable factors that we expected to affect detection. Accordingly, we consider our estimates to represent relative abundance, acknowledging that they are likely biased. Notably, if violations of assumptions vary across space (e.g., different degrees of closure at different sites) estimates may not provide a valid index of abundance. While there is likely some variation in bias among sites, based on observations during surveys, we believe such variation is likely negligible.

Additional limitations of this unmarked survey method include that relative abundance estimates were sensitive to timing of sampling within the active season or location of sampling along a given creek. This is evidenced by inconsistencies identified in the assigned abundance category of a creek when sampling the same site on multiple dates and when sampling different sites along the same creek on the same date (e.g., Fig. 4). The effect of day of year on mean expected abundance (Fig. 3E) also corroborates that relative abundance is affected by timing. Thus, when and where sampling is conducted could affect management decisions.

We also assessed whether estimates were robust to amount of survey effort. The large range in tidal amplitude experienced in Georgia salt marshes during semidiurnal tide cycles creates a short (∼ 6 hr) time window in which the water line is below the high marsh, limiting the number of sites that can be surveyed in a single day. Because sampling many sites is important to meet the monitoring objectives of the DNR, sampling efficiency is a priority. We found that as effort increased from 3 to 10 repeated counts of the site, the number of sites that were assigned to the higher abundance categories also increased. During our surveys, each scan of a site took an average of 69 sec (range = 10–213 sec/scan), with duration primarily affected by the size of the sample area, but also impacted by other factors (e.g., waves). Thus, on average, with 1 min allotted between scans, a 3-scan survey takes 5.5 min, a 5-scan survey takes 9.8 min, and a 10-scan survey takes 20.5 min. This information can be useful in weighing the importance of conducting surveys over fewer days versus the importance of being able to confidently assign creeks to higher abundance categories, allowing less stringent actions at more locations.

Model selection indicated strong support for effects of salinity, distance to potential nesting habitat, and the day of year on relative abundance, but little support for the effects of known crabbing activity or proportion of shoreline armoring. The quadratic relationship observed between day of year and abundance is likely an artifact of nesting phenology and sampling design. Because of limited boat access, most surveys were conducted at sites that were accessible by land (n = 542 surveys by land vs. 64 surveys by boat)—including bridges, fishing piers, public and private docks, uplands, and a few high marsh locations where the marsh was safely accessed by foot. Many land-accessible sites were on or adjacent to potential nesting habitat and often terrapins observed from these locations appeared to be females waiting for suitable conditions to nest (Feinberg 2004; Butler et al. 2018). Conversely, most of the surveys conducted by boat were in more remote marsh locations (i.e., foraging habitat). Expected abundance peaked in late May, aligning with what we could expect for terrapin nesting activity in the region (Crawford et al. 2014b). Levasseur et al. (2019) similarly found that abundance varied seasonally in Wellfleet Bay, Massachusetts, consistent with phenological and spatiotemporal expectations for mating and nesting activity in the north (Levasseur et al. 2019, 2023). If future surveys are conducted by boat to target terrapins in their foraging habitat, as planned by GA DNR, we expect that the direction of the day-of-year versus abundance relationship will change, with abundance likely greatest early in the season (i.e., prenesting).

The positive relationship observed between expected abundance and distance to potential nesting habitat reflects the larger numbers of terrapins typically observed at sites accessed by boat (mean heads/scan = 0.61, range = 0–17) compared to sites accessed by land (mean heads/scan = 0.25, range = 0–9). The candidate model set we considered limited our ability to determine the biological mechanisms driving this relationship, but there are likely multiple contributing factors, including the nesting phenology of terrapins described above and the range of dates sampled, which included times outside of peak nesting season. Additionally, sites accessed by land included locations with higher human activity (e.g., boating, fishing) and locations near more development (e.g., creeks passing under or near roads)—factors that could affect habitat quality. Whether or not we observed crabbing activity in the sections of creek visible to us from our land-access survey points, these less-remote creeks also may be crabbed more heavily because of their convenient access. Future work could assess support for a model including an interaction term between distance to potential nesting habitat and day of year, if there are sufficient data to support the number of model parameters, to improve interpretation of the expected abundance and distance to potential nesting habitat relationship.

Counterintuitive in view of previous research (Harden and Southwood Williard 2018), we also observed that abundance increased with the average salinity of a creek. While terrapins are adapted to brackish water environments, substantial energy is required for physiological osmoregulation and terrapins have been found to eat less and grow slower when in high-salinity environments (Dunson 1985; Davenport and Ward 1993; Holliday et al. 2009; Southwood Williard et al. 2019; Ashley et al. 2021). Despite this, when other covariates were held at mean values, model predictions yielded an expected abundance of < 1 individual at salinities ≤ 30 ppt and > 1 individual at salinities ≥ 31 ppt. Additionally, we never observed terrapins at the 5 creeks with the lowest mean salinity records (range = 0.11–3.77 ppt) that otherwise appeared to be suitable habitat. Although the reason for this relationship is unknown, it may be important to control for the effect of salinity on abundance when making inferences about causal effects of suspected threats to terrapin populations.

Our inability to detect negative effects of crabbing or shoreline armoring on abundance differs from previous studies (Dorcas et al. 2007; Grosse et al. 2011; Isdell et al. 2015; Lovich et al. 2018). The measurement of crabbing activity used in our study was very coarse compared to these other studies. We could not account for amounts of crabbing effort (i.e., number of pots and duration of crabbing), locations of crab pots, and soak times which are factors that could influence the effect of crabbing on abundance. It is also possible that some creeks with no observed crabbing were being crabbed within areas not visible to us from our terrestrial survey points. Crabbers in Georgia are also known to move the location of their crab pots during the year, so it is possible that our surveys might not have coincided with crabbing at some sites. The lack of support for an effect of shoreline armoring may have been affected by the scale we considered (within 100 m of the sample area), which was smaller than the 1-km scale at which armoring was found to affect terrapins in Virginia (Isdell et al. 2015). It is also possible that our study sites did not have enough shoreline armoring to affect terrapin abundance. Among our study sites, the distribution of proportion of shoreline armoring at the 1-km spatial scale was highly skewed with 90% of values < 0.11 (range = 0–0.77), so we did not consider this scale in our analyses. Isdell et al. (2015) stratified sampling to survey a gradient of shoreline armoring and identified a threshold of 0.17 armored shore to be important at the 1-km scale. Therefore, our study sites appear to have generally less shoreline armoring than the study sites in Virginia.

Comparisons between this study and sites with naturally high terrapin densities (i.e., creeks without crabbing pressure) are difficult to make. No previous studies have used the same method to estimate relative abundance in the tidal creek environment; thus, comparisons should be interpreted cautiously. After rescaling estimates from the largest spatial-scale capture−recapture study in Georgia to match our scale (individuals per 6000 m²), we found that only 1 of our sites during a single survey exceeded the lower 95% confidence boundary for estimated abundance at sites with no crabbing pressure (Grosse et al 2011). Additionally, among 13 head count creeks that were also sampled by Grosse et al (2011), estimates using the head count method were consistently lower than estimates from seining. These studies were conducted 14 years apart, and it is possible that abundance truly decreased within all 13 creeks. However, N-mixture model estimates have been found to be biased low when detection is low, while capture–recapture model estimates were not (Couturier 2013). Because terrapins spend time below the surface of the water, may be hidden in vegetation within the high marsh, and can be otherwise difficult to observe, low detectability is a known issue. We also found that most of our surveys resulted in estimates in the < 1 individuals per 6000 m² category. These low abundance estimates could be, in part, related to the sensitivity of estimates to location along the creek, and may not characterize abundance of the entire creek. The percentage of creeks estimated by Grosse et al. (2011) to have ≥ 1 terrapin (86%) vs. the percentage of surveys in the current study with an estimate of ≥ 1 terrapin (30%), also suggests that the head count method may underestimate the number of occupied creeks.

Because relative abundance estimates using head count methods are sensitive to timing and location of sampling, we recommend standardizing these variables across years. Although we sampled some creeks by boat, locations of sampling during our surveys were primarily dictated by accessibility by land so we could not control the locations of sampling along most creeks surveyed. It is not surprising that terrapin abundance varies along the length of a creek based on our previous observations during capture–recapture surveys, during which some sections of creek are observed to consistently have relatively more or less terrapins across sampling years. For sampling conducted by boat, locations could be selected to be more representative of terrapin abundance within the creek. While accessing creeks for head count surveys and other monitoring efforts, we observed that terrapins typically tended to be most abundant within the terminal ends of the creeks. Therefore, this location may best characterize the level of terrapin abundance within the creek for the monitoring objectives of GA DNR.

The limited amount of data collected by boat in this study restricted our ability to estimate optimal survey timing for monitoring terrapins in their foraging habitat. However, based on our knowledge of terrapin nesting phenology, conducting surveys within a short timeframe early in the season when females are most likely to be found within the medium and smaller tidal creeks potentially farther from nesting areas may be ideal. Future work to inform optimal monitoring should use a larger sample size of boat surveys to estimate optimal survey timing and assess the number of annual surveys and the number of sample locations along a creek needed to reduce the risk of mis-categorizing a creek to an acceptable level.

Finally, integrating head count data with capture–recapture data should be explored as a means of improving estimates. Integrated models can improve accuracy and precision of estimates by jointly estimating parameters shared by more than 1 dataset (Zipkin and Saunders 2018; Frost et al. 2023). For example, in a study of western bluebirds (Sialia mexicana), Sanderlin et al. (2019) found that integrating presence-absence data collected at 149 sites with capture–recapture data collected at a subset of the same sites (n = 72) improved estimate precision, relative to modeling either of the datasets separately. A similar approach could be evaluated for diamondback terrapins in Georgia to estimate precision gains expected with integrating capture–recapture data collected at a subset of head count locations.

Conducting surveys by land accessible points may still be useful under some circumstances if there is an appropriate monitoring objective. This method is easy to learn and could harness volunteer effort to conduct many surveys over a short period, similar to programs implemented for other taxa. For birds, the Christmas Bird Count and Breeding Bird Survey have demonstrated how large amounts of volunteer-collected count data, while imperfect, can be used to monitor large-scale population trends and address other ecological questions (Bock and Root 1981; Hagan 1993; Sauer et al. 2003; Silvertown 2009). Volunteer-collected counts can also be combined with other data types via integrated models to improve reliability and inferential capacity (Hanks et al. 2011; Steger et al. 2017; Robinson et al. 2018; Sun et al. 2019).

Diamondback terrapins are challenging to monitor because of their difficult-to-access habitat and daily and seasonal variation in habitat use. On a daily basis, tide cycles affect whether terrapins are likely to be found in creeks or the high marsh, which is virtually inaccessible for monitoring. Throughout the active season, nesting migrations affect the location of mature females. Despite these challenges, monitoring is necessary to understand the status of populations and effectively apply targeted management. Capture–recapture studies of terrapins within their aquatic habitat are not feasible to conduct at broad spatial scales long term. Therefore, if used judiciously, a method using head count surveys and N-mixture models may be the most effective means of monitoring state-wide.

This study can serve as a starting point for refining this method to meet the needs of the Georgia DNR. Because terrapin distributions within creeks tend to be clustered, we recommend surveying nonrandomly within each creek to monitor sites that best characterize abundance. Conducting surveys by boat is likely necessary to target such locations. We also recommend conducting surveys over short and consistent time periods within the active season. Prenesting season is likely ideal, but this should be evaluated with more boat survey data. It may also be useful to consider sampling plans where creeks that meet certain conditions are resampled. For example, if a creek shifts to a lower abundance category across years, it could be resampled at additional times (and possibly additional locations) to confirm the change in abundance and need for restrictive management actions. Follow-up studies using these recommendations could also help determine whether relative abundance categories should be revised from those used in this study to more effectively inform management decisions. Other states with similar salt marsh tidal creek habitat may benefit from using this same sampling approach if objectives are the same. However, if habitat, sampling conditions, or monitoring objectives differ, we recommend reevaluating the methodology under the new context.