Multi-year temporal and seasonal abundance patterns of common bottlenose dolphins near Sarasota, Florida

1 Introduction

Population monitoring is a fundamental tool for understanding wildlife ecology and population dynamics and for establishing and evaluating the effectiveness of conservation measures and policies (Moussy et al., 2022). Data on long-term trends in abundance and survival are essential for predicting population viability and evaluating the stability and health of a population over time, especially for species with slow life histories (Gould et al., 2024). Updated abundance estimates at regular intervals are also required for conservation efforts in which resource managers calculate allowable anthropogenically caused mortalities. For marine mammals in the U.S., this is exemplified by calculating the annual potential biological removal (PBR) level allowable for each designated marine mammal management unit, referred to as a stock (Wade and Angliss, 1997). Unfortunately, obtaining accurate abundance estimates can be challenging. This is especially true for marine species that have the capacity for long-distance movements, are not easily caught or tagged, and are not readily visible at the surface of the water or from above via aerial surveys.

For common bottlenose dolphins (Tursiops truncatus), abundance estimation is further complicated by a complex social structure which makes it challenging to identify biologically meaningful population units (Wells and Scott, 2024). Bottlenose dolphins (hereafter dolphins) inhabiting the bay, sound, and estuary waters near Sarasota, Florida have been studied by the Sarasota Dolphin Research Program (SDRP) since 1970 (Irvine et al., 1981; Wells, 2009, 2014, 2020). Early studies revealed dolphins exhibiting year-round fidelity to the inshore system (now referred to as a resident community; Irvine and Wells, 1972; Wells et al., 1980, 1987, 1999; Wells, 2014). The resident Sarasota dolphin community overlaps spatially with non-resident individuals and groups that either pass through or have overlapping ranges, but are socially distinct (Fazioli et al., 2006; Urian et al., 2009; Wells et al., 1987; Wells, 2014; Wells and McHugh, 2025) and genetically distinguishable (Duffield and Wells, 1986, 2023; Sellas et al., 2005; Wells, 2014). Most dolphins in the resident community are individually identifiable due to natural markings, tag scars, or freeze-brands applied during catch-and-release studies (Wells, 2009). These distinctive features in combination with a high level of field survey effort facilitate abundance estimates from direct censuses using photographic-identification (photo-ID) to track individuals visually (Wells and Scott, 1990; Wells, 2009; Wells et al., 2025). This approach also avoids issues associated with needing to account for unknown non-residents that also utilize the inshore system. In general, the number of Sarasota dolphin community members identified through direct census has increased since the 1995 ban of fishing entanglement nets in inshore Florida waters (Wells et al., 2025). However, results from population viability modeling have identified a recent slowing of this trend, suggesting that the community may be reaching carrying capacity (Lacy et al., 2021; Wells et al., 2025). Unfortunately, given the substantial time and resource demands required for direct census, collecting comparable data elsewhere is not often feasible, leading to difficulties comparing and ensuring consistency in abundance estimates across different bay systems in the southeastern U.S (Rosel et al., 2011; Urian et al., 2015).

As technologies have advanced, so too have modeling capabilities. As of 2011, the recommended approach for estimating the abundance of bottlenose dolphins in inshore habitats was to conduct standardized photo-ID (Würsig and Würsig, 1977; Würsig and Jefferson, 1990) capture-mark-recapture (CMR) surveys following Pollock’s robust design sampling strategy (e.g., seasonally spaced primary sampling periods comprised of multiple, quick secondary sessions; Pollock, 1982) and to screen images based on a standard set of best practices (Melancon et al., 2011; Rosel et al., 2011; Urian et al., 1999, 2013, 2015). These recommendations were designed to maximize capture probabilities, account for heterogeneity in capture probabilities due to the inclusion of non-residents in areas where the residents were unknown, combat violations of CMR assumptions, deal with temporary emigration, and account for unmarked individuals in the population. Though abundance estimates from direct census may be more accurate for a given resident community that represents a biologically meaningful unit, the robust design approach to survey design and corresponding analyses allow for comparable results across studies and through time.

Only one set of CMR abundance estimates have been reported for dolphins inhabiting the Sarasota Bay area using the updated recommendations (Tyson Moore and Wells, 2016). However, it was restricted to data obtained across four seasons from a single year, limiting the analytical power needed to model demographic processes. It also didn’t account for non-residents, so it’s unclear how well those results represented the well-described resident community. Here, we estimate annual abundance of bottlenose dolphins inhabiting the inshore Sarasota and Little Sarasota Bays over a ten-year period (2013-2022) using an approach comparable to that used by others in the southeastern U.S. where less is known about the dolphins. First, we investigated the existence of heterogeneity of capture probabilities and transience effects in the dataset by developing and testing appropriate models. We then used Pollock’s robust design models to derive estimates of population size, survival, and temporary emigration rates, taking transience effects into account. To assess how our model output aligned with traditional methods, we compared results to recently published minimum abundance estimates for the Sarasota resident dolphin community derived from direct census (Wells et al., 2025). If results obtained from our model using Pollock’s robust design do indeed capture abundance based primarily on the resident community, then we would expect little variation in modeled abundance estimates from year to year over the past decade, supporting conclusions from census data and adding support to the observation that the Sarasota dolphin community may be reaching carrying capacity (Lacy et al., 2021). Abundance estimate results will be provided to NOAA’s National Marine Fisheries Service (NMFS)—the federal agency charged with managing dolphin stocks—to update the Stock Assessment Report, as required by the Marine Mammal Protection Act.

2 Materials and methods

2.1 Study area

Sarasota Bay (27.3460°N, 82.5673°W) is located on the central Gulf coast of Florida, USA and connects Little Sarasota Bay to the south and Tampa Bay to the north. Since mid-1992, standardized photo-ID CMR surveys have been conducted each month in and around Sarasota Bay (Wells, 2009, 2014) ranging from roughly south Tampa Bay to Venice Inlet (approximately 245 km²) (Figure 1). Though dolphins from the resident community are known to range into the adjacent nearshore Gulf of Mexico within several kilometers of shore, an estimated 89% of sightings of dolphins considered to be Sarasota Bay residents are from shallow (<4 m deep), inshore estuarine habitats (Wells, 2014). Therefore, in order to decrease the chances of erroneously including non-residents in analyses, the study area here is limited to waters inside the barrier islands and in passes, coinciding with NOAA’s Bay, Sound and Estuary Bottlenose Dolphin Stock Block 25 designated by Hayes et al. (2023) (approximately 149 km²; Figure 1A).

2.2 Data collection

Since 2013, quarterly synoptic surveys have been conducted once per season (i.e., four primary periods per year). To follow Pollock’s robust design, the entire system was covered by three boats across three routes (north, middle, and south) (Figure 1B) operating on the same day, and that effort was repeated for three days (i.e., three secondary sampling sessions). These synoptic surveys were conducted in as short a time as possible, often over three consecutive days whenever weather conditions allowed. Between primary periods, dolphin populations were open to births, deaths, immigration, and emigration. Within primary periods, populations were assumed to be closed. In order to remain comparable with other studies, only seasonal synoptic data were included for analyses from 2013-2022, resulting in a dataset with 40 primary periods and 120 secondary sessions over 10 years. During this time period, winter surveys were conducted in January or February (with the exception of 2015 when surveys were delayed until March due to poor weather conditions), spring surveys in April or May, summer surveys in July or August, and fall surveys in October or November (Table 1).

Table 1. Summary of primary sampling periods; each survey day represents a secondary sampling session in which three boats covered all regions of the primary study area in a single day (synoptics effort).

Seasons were defined as Winter: Dec-Feb; Spring: Mar-May; Summer: June-Aug; Fall: Sept-Nov; *Conducted winter surveys in March this year.

Sampling duration was defined as the number of calendar days between the start and end of the surveys, during which the system was covered three times.

Surveys were conducted from 6–7 m, outboard-powered, center-console research vessels that traveled at the minimum speed necessary to maintain a plane during good sighting conditions (i.e., 14–18 knots; Beaufort Sea State ≤ 3) (Rosel et al., 2011). At least three observers were on board each vessel with the exception of during the global pandemic (spring 2020 through spring 2021); only two observers were on each vessel to minimize COVID-19 exposure risk. When a group of dolphins was sighted, the boat slowly approached the animals and recorded their GPS position, along with other environmental and biological data. A Nikon DSLR camera with 70 to 300 mm lens was used to take photographs of the dorsal fin and other distinguishing features (e.g., peduncle) of every dolphin in the group. In larger groups, an attempt to obtain photographs of as many of the animals present was made, across both marked and unmarked individuals. We defined a dolphin group as all dolphins in relatively close proximity (<100 m), engaged in similar behavior, and generally heading in the same direction (Wells et al., 1980).

Photo-ID processing and analyses followed established methods (Melancon et al., 2011; Rosel et al., 2011; Urian et al., 1999, 2013, 2015). Briefly, dorsal-fin images were cropped and graded independently for photographic quality (Q-1: excellent quality; Q-2: average quality; Q-3: poor quality) and overall distinctiveness (D-1: high/very distinctive; D-2: average distinctiveness; D-3: low distinctiveness) of the fin based on an established point method (Urian et al., 2013). To minimize false positive matches, as many distinctive features as possible were used to confirm matches. A second investigator verified every potential match of an individual dolphin from one encounter to another and double-checked the entire catalog for potential matches or mismatches that may have been overlooked. Images of newly identified distinctive dolphins and their respective sighting data (e.g., time, location, identification code) were catalogued using FinBase, a customized Microsoft Access-based database (Adams et al., 2006; Melancon et al., 2011).

2.3 Analyses

In order to minimize subjectivity and meet the assumptions of the robust design model in analyses, sighting histories were screened to include distinct (D-1 or D-2 only), non-calf dolphins from excellent or good images (Q-1 or Q-2). These individuals were considered to be marked and all others unmarked—only marked animals were used in analyses. Here, “marked” refers to dolphins with sufficiently distinctive features for reliable photo-identification, regardless of whether those features were naturally acquired or resulted from prior research activities (e.g., tag scars or freeze-brands); all captures and recaptures in this study were based on photographic identification. Unmarked animals were incorporated in post hoc calculations to adjust derived abundance estimates based on the proportion of marked animals in the study (Friday et al., 2008; Rosel et al., 2011; Stevick et al., 2001; Urian et al., 2013). Calves are defined in the field as an identifiable individual known to be a dependent calf, or smaller individual consistently swimming alongside an adult in a stereotypic position. Calves are confirmed when seen with a female presumed to be its mother in at least three consecutive sightings. Relationships were typically confirmed genetically from samples collected during catch-and-release health assessments or biopsy dart sampling (Duffield and Wells, 1991, 2023). The total number of dolphins identified during the study period and the rate of identification were plotted using a discovery curve, which visualizes the accumulation of newly identified individuals through time and provides insight into photo-ID catalog growth, reflecting the degree of population openness over the study period and processes such as recruitment and immigration (Wilson et al., 1999).

2.3.1 Testing model assumptions

Common challenges in modeling dolphin population parameters that can lead to violations of model assumptions include accounting for heterogeneity in capture probabilities and transience effects (e.g., violation of the assumption that individuals newly marked have the same apparent survival as previously marked individuals) (Rosel et al., 2011; Urian et al., 2015). Unmodeled heterogeneity of capture probability can negatively bias abundance estimates because individuals with low capture probabilities are underrepresented in the observed sample (Pledger et al., 2010). Excluding poorly marked individuals from analyses helps address this issue by reducing heterogeneity associated with mark distinctiveness. However, inclusion of non-resident dolphins introduces another source of heterogeneity that can influence multiple parameters. Because non-residents are less likely to be encountered repeatedly, their inclusion may depress estimated recapture probabilities and potentially inflate abundance estimates. Permanent departure by non-residents may also complicate estimation of γ (defined below) because it can be mistaken for temporary absence, causing estimated temporary emigration rates to differ from true temporary emigration rates. Finally, animals that pass through a study area and are never seen again are treated as if they died, resulting in a negative bias in apparent survival rates (Pradel et al., 1997). Therefore, we first ran exploratory models to test violations of model assumptions and inform model development.

There are few goodness-of-fit (GOF) tests suitable for robust design models and none test for heterogeneity in capture probabilities (Laake, 2013; White and Burnham, 1999); however, the piece of the robust design model that assumes a closed population within primary periods makes the same assumptions as those for a closed population using a Cormack-Jolly-Seber (CJS) CMR model. Therefore, the capture history data were first formatted to fit a CJS model, which was run using program MARK (White and Burnham, 1999) version 11.0 accessed via the RMark package (Laake, 2013) in R v4.6.0 (R Core Team, 2025) using RStudio v2026.4.0.0 (Posit team, 2026). The R2ucare R package (Gimenez et al., 2018) was used to test for transience effects (i.e., to test if the presence of non-residents in our dataset was high enough to violate model assumptions) and trap dependence (i.e., test if the probability of an animal being captured was influenced by its previous capture history). Trap dependence typically occurs in species that are physically caught and recaptured, after which they become trap-shy (or trap-happy), violating the assumption that all animals have the same probability of recapture. In bottlenose dolphin research, “marking” individuals using non-invasive photography is not expected to result in behavioral responses that would impact the probability of recapture, in which case, trap dependence is not an issue. However, a significant GOF test for trap dependence may result from other sources of heterogeneity in capture probabilities, such as temporary emigration (e.g., some residents were not available at the time of sampling) or the presence of non-resident dolphins, among others (Jeyam et al., 2018), in which case, this test can still be informative. Model GOF was poor (χ² (118) = 622.5, p < 0.01) and component tests consistent with trap dependence (χ² (37) = 160.5) and transience effects (χ² (30) = 175.1) were significant (p < 0.01). After removing the significant components from the overall CJS statistic (using the approach suggested by Gimenez et al., 2018), the remaining component was not significant (χ² (37) = 50.9; p = 0.06), indicating that little unexplained lack of fit remained after accounting for these components. Additionally, when identical robust design models were run with parameters accounting for heterogeneity (“RDFullHet” in RMARK) and without (“Robust” in RMARK)—as suggested by Jeyam et al. (2018) and Pledger et al. (2010)—the models accounting for heterogeneity consistently outperformed those without (i.e., had lower AIC values). These diagnostics informed the inclusion of transience effects, finite-mixture detection heterogeneity, and temporary emigration in the robust-design candidate models.

Biases associated with heterogeneity of capture probabilities can be mitigated directly in a robust design model. However, accounting for transience effects without assigning specific individuals to residency categories is less straightforward. We chose to exclude obvious transients from our dataset to reduce heterogeneity and align our approach with research in the southeastern U.S., where datasets permitting individual-based residency assignments are uncommon. In studies with limited knowledge of community structure, animals seen only once during the study or in only one primary period (“single sightings”) are often presumed to be transient. Because our analyses included only synoptic surveys conducted within the core area (see Figure 1A for core vs. total area), some animals likely appeared to be “single sightings” despite having additional records from other survey types or elsewhere in the total study area. To avoid incorrectly excluding these individuals, we reviewed their complete ten-year sighting histories across all survey areas and survey efforts. This review allowed us to distinguish obvious transients that rarely used the core study area (and could be excluded) from animals that had additional sightings elsewhere or during other survey efforts. The latter could include animals from the resident community (e.g., animals first entering the dataset after aging out of the calf stage late in the study) but was more likely to include members of adjacent communities (i.e., presumed non-residents) whose use of the core area was too frequent for them to be classified as true transients. We excluded individuals from the analyses if they met any of the following criteria when all sighting data from the ten-year study period—not only synoptic surveys—were considered:

• only seen in the core study area in a single primary period but never again in the ten-year study (may have been resighted outside of the core area).
• seen in the core study area more than once within a single year but never again in the ten-year study period (inside or outside the core area).
• seen in the core study area two times or fewer, but the majority of their sightings were limited to passes and/or along the barrier island coastline.

Individuals with single sightings were retained if they were seen in the study area during other survey types (e.g., monthly surveys) and thus were simply flagged due to our initial screening criteria as identified above. Individuals were also retained if there was a record of a valid death precluding us from making decisions about its habitat use. After obvious transients were removed (see results), GOF tests were re-run, but the results were unchanged, suggesting that screening alone did not eliminate encounter-history patterns associated with transience effects and residual heterogeneity. This was consistent with the retention of non-resident animals, such as individuals from adjacent communities. Similarly, “RDFullHet” continued to outperform otherwise equivalent “Robust” models that did not account for capture heterogeneity. We therefore incorporated finite-mixture capture heterogeneity and an “age-dependent” (a.k.a. time-since-marking) transience effect into the model design as explained below. Here, “transience effect” refers to a statistical pattern and does not imply that all retained animals contributing to this pattern were biologically transient.

2.3.2 Abundance

The following parameters were estimated using the Robust Design with Full Heterogeneity Estimator (RDFullHet) in RMARK:

• apparent survival (ϕ) - the probability of surviving from occasion t to t+1.
• recapture probability (p_ts) - the probability of capturing an individual in secondary session (s) of primary period (t), given that it is alive and in the sampled area during period t.
• discrete mixture term (pi) - the probability that an individual occurs in mixture A.
• gamma double prime (γ”) - the probability of an individual being unavailable for capture in a given primary period (t) given it was present in the previous primary session (t-1) and survived (i.e., an individual temporarily emigrated from the study area between sampling occasions).
• gamma prime (γ’) - the probability of an individual being unavailable for capture at time (t) given that it was not present in the previous primary session (t-1), survived, and remains unavailable.
• number of unmarked individuals (f0) - estimate of the number that were never encountered.
• derived abundance (Nhat) - estimated population size based on marked animals only.

For the CMR analysis, a variety of candidate models were fitted. Covariates for sampling effort were not included given the consistency in effort over time (Tables 1, 2). Heterogeneity in capture probabilities was accounted for by including a constant mixture term (pi), in which p (initial capture probability) was allowed to vary by mixture group (m = 2 groups based on recommendations by Pledger, 2000) for all models. As mentioned above, behavioral responses to photographic “marking” of animals were not expected to further contribute to heterogeneity; thus recapture probability (c) was set equal to initial capture probability (p). Preliminary candidate models with no temporary emigration (γ” = γ’ = 0) were used to assess the additive effects of heterogeneity on capture probability due to: year (10 levels); season (4 levels); time (a factor variable with 40 levels, one for each primary period); year + season; and year + season + time (heterogeneity was not modeled by secondary session given that surveys were conducted consecutively within each primary period). Modeling capture probabilities as heterogeneity + year + season + time resulted in overparameterized models with non-identifiable capture probability terms. Models including that covariate combination were dropped from the candidate model set. Transience effects were accounted for using an “age-dependent” modeling approach whereby survival probability was allowed to differ over the first interval after first mark compared to that of subsequent intervals using a “transience” covariate (Pradel et al., 1997; implemented by Silva et al., 2009). Candidate models with no temporary emigration (γ” = γ’ = 0) were also used to assess the effects of the same year, season, and time (i.e., primary period) covariates on survival probability, as well as the interactions between transience and time or year. When survival was estimated by primary period (t), candidate models either did not converge or were overparameterized, resulting in non-identifiable survival terms. Using starting values informed from an initial model or simulated annealing techniques (see Goffe et al., 1994) did not resolve the issues. All models in which survival varied by primary period were therefore removed from the final set of candidate models. Remaining survival terms (listed in Table 3) were modeled using a sin link function to pull them away from an upper boundary of 1, which allowed for the standard error to be calculated (Gimenez et al., 2009). To estimate temporary emigration between primary periods, both constant and time-varying Markovian movement terms were incorporated. Preliminary model runs revealed a lack of estimability for gamma parameters, possibly due to transience effects. Adding a transience covariate improved estimates but resulted in estimates near a lower boundary of 0. Therefore, gamma parameters were also modeled using a sin link function. Random emigration models were not included since it was already known that the probability of leaving the study area and returning was not random.

Table 2. Photographic identification effort for each year during the primary periods (n=40), including the total number of dolphins, calves and neonates estimated in the field and identified during post-processing.

2015 winter survey was conducted in March. Individual dolphins may have been seen during multiple sightings and/or on multiple days.

Table 3. Covariate combinations included for modeling apparent survival, capture probability, and emigration parameters under Markovian emigration.

Covariate combinations included for modeling apparent survival (ϕ), capture probability (p), gamma double prime (γ”), and gamma prime (γ’) under Markovian emigration (γ”(x) ≠ γ’(x)); the constant dot model is indicated by 0 and (.); time (t); transience (A).1 denotes intercept-only model (i.e., with no covariates). Covariates are defined in the text.

The final set of covariates included are listed in Table 3. All models were fitted using an initial model to provide starting values based on the most complex model. Markovian emigration was modeled as (γ”(x) ≠ γ’(x)), where x was either constant (dot model), time-varying (t), included a transience covariate (A), or included an interaction term between transience and time (At), resulting in 8 possible combinations. Since the transience covariate was structured to account for animals only seen at capture and never again, it was modeled for γ” only, as γ’ is not estimated over the first interval after capture in robust design models. For models in which survival probability varied over any version of time (e.g., year or season), gamma parameters were constrained over the last two sampling periods (i.e., γk’ = γk’ − 1 and γk” = γk” − 1, where k is the last primary period) to prevent them from confounding with the last two survival parameter estimates (Kendall et al., 1997).

There is no GOF test available for the fitted robust-design model (White and Burnham, 1999); therefore, the variance inflation factor (ĉ) could not be estimated empirically. Models were initially compared using Akaike’s Information Criterion with small-sample correction (AICc) (Burnham and Anderson, 1992; Burnham et al., 2011), corresponding to ĉ = 1. To evaluate the robustness of model selection results to potential overdispersion, model rankings were recalculated using QAICc under assumed values of ĉ = 2, 3, and 4. Additional analyses at 0.5-unit increments were used to identify the approximate ĉ value at which the top-ranked model or supported capture-probability structure first changed (only whole-number values of ĉ are reported). For each assumed value, models were compared using QAICc, ΔQAICc, QAICc weights, and QDeviance. Because model selection changed with increasing ĉ, model averaging was repeated under moderate (ĉ = 2) and high (ĉ = 4) assumed overdispersion. For each analysis, model averaging used the smallest set of top-ranked models accounting for approximately 98% of cumulative model weight (three models for the original analysis and ĉ = 2, and seven models for ĉ = 4). Model-averaged estimates, unconditional standard errors, and 95% confidence intervals were compared among analyses. Because the values of ĉ were assumed rather than estimated, these analyses were interpreted as sensitivity tests rather than estimates of the actual degree of overdispersion. Derived abundance estimates were based on the marked proportion of animals only. An adjustment based on the proportion of unmarked animals was applied to estimate total population size (Hammond et al., 1990; Williams et al., 2002) based on the following calculation:

Ntotal=Ňm/θ

where N_total is the estimated total population size, Ň_m is the estimated distinctly marked population size, and θ is the proportion of distinctive dolphins in the population, including calves (Williams et al., 2002; Wilson et al., 1999). Variance and standard errors of N_total were calculated according to the delta method (Williams et al., 2002) and the log-normal 95% confidence intervals were calculated using formulas provided by Burnham et al. (1987).

3 Results

3.1 Survey summary

Seasonal surveys conducted over the ten-year study period (2013-2022), resulted in a total of 120 sampling days (secondary sessions) over a 3,529-day period for a sampling frequency (m) of 0.03 (Morteo et al., 2012). An average of 257.7 (± 24.2) hours were spent on the water per year, covering an average of 4,147.9 (± 35.0) km per year (Table 2). Surveys resulted in 2,389 dolphin group sightings (encounters) in the core study area, with 493 unique individuals being identified across all distinctiveness classifications (D1-D4), life stages, and before screening for transience and photo quality. After screening for photo quality (Q1 and Q2; removed 13% (n = 64) of individuals) and distinctiveness (D1 and D2; removed additional 25.8% (n =127) of individuals), and removing animals that remained in the calf/neonate life stage during the time duration of the analysis (n=12), 290 remaining non-calf individuals were identified, 83 (28.6%) of which were only seen once across 40 primary periods (but could have been seen more than once within a primary period). Following the transience screening rules identified above, 40 of the 83 animals were removed, resulting in 250 unique non-calf dolphins (43, or 17.2%, of those being animals with single sightings) included for analysis. A given unique individual was sighted in an average of 9.19 primary periods (median [max] ± SD: 5 [32] ± 8.23).

The total number of uniquely identifiable and marked non-calves added to the catalog increased steadily over the study period, but the rate of increase fell following the first two years (2013–2014) of sampling (or first eight primary periods) (Figure 2). After 2014, the number of unique newly marked non-calves observed per sampling period remained relatively stable (range 0-11). The sampling periods with the highest numbers of newly marked non-calves (after 2014) observed were the spring of 2020 (primary period 30; n = 11), fall of 2019 (primary period 28; n = 9), and summer of 2017 (primary period 19; n = 8). The summer of 2017 was also the sampling period with the highest number of total sightings (tied with fall of 2019 at 83 sightings; Table 2), highest number of total dolphins sighted in a single period (n = 454) and highest number of marked dolphins in a single period (n = 334). The rate of increase for new animals added to the catalog slowed again in 2020 (primary period 30).

3.2 Estimates from robust design

The final set of candidate models included 175 robust design CMR models. Results from the top 10 (lowest AIC_c) are shown in Table 4. The best fit model (AIC_c = 1231.93) only fit 1.2x as well as the next candidate model. The next two best fit models had matching emigration and capture probability terms but varied in survival probability parameters. The best fit model only accounted for ~43% of the support in the data. The top two models accounted for 81% and the top three accounted for 97% of the support in the data. Model selection was sensitive to increasing assumed overdispersion (Supplementary Table S1). At ĉ = 2, the rank of the original top two models (S.season and S.A, respectfully) switched, and the original third-ranking model (S.Ayear) was replaced by a null model for survival (S.0). The structure of the remaining parameters remained unchanged. At ĉ = 3 and 4, QAICc favored substantially simpler capture-probability structures, with the top-ranked models allowing capture probability to vary only by mixture group (p.m).

Table 4. The 10 CMR models fitted with the lowest AIC_c.

MA00 refers to Markovian movement (M) in which γ” (A0) was modeled with a transience term (A) but not by time (0) and γ’ was estimated as a dot model (0). npar is the number of parameters; deltaAIC_c is the AIC_c difference between the model and that with the lowest AIC_c; deviance quantifies the discrepancy between the fitted model and the saturated model; transience term (A); intercept-only model (1); Markovian emigration (M); dot model (0); time (t) varies by primary period; probability of detection (p) varies by mixture group (m) and primary period (s).

Under the original AICc analysis, all three top models supported a Markovian emigration process in which the probability of temporary emigration depended on an individual’s previous availability state. Specifically, the probability of becoming unavailable (γ″) was explained by the transience term (A) and the probability of remaining unavailable (γ′) was constant (γ′ = 0.89). As expected, the model-averaged probability of becoming unavailable (γ″) was high for animals seen only once (presumed non-residents; some were resighted during monthly surveys or outside the core area: γ” = 0.32; SE = 0.06, 95% CI 0.22 – 0.43) compared to those seen over subsequent intervals (presumed residents: γ” = 0.05; SE = 0.01, 95% CI 0.03 – 0.07; Figure 3), suggesting little movement out of the core study area between primary periods by presumed residents. Despite changes in model rankings and weights, model-averaged temporary emigration estimates remained stable when ĉ > 1 (Supplementary Table S2). Under ĉ = 4, the probability of becoming unavailable was 0.049 for presumed residents (SE = 0.02, 95% CI 0.03–0.09) and 0.30 for presumed non-residents (SE = 0.10, 95% CI 0.15–0.52), while the probability of remaining unavailable was 0.90 (SE = 0.04, 95% CI 0.80–0.95). Relative to the original analysis, these point estimates changed by no more than 0.014, whereas their standard errors increased by approximately 77–97%.

Across 40 primary periods (2013–2022), the number of unique dolphins recorded using the core area varied from a low of 34 individuals in both the summer of 2015 and the spring of 2016 to a maximum of 91 individuals in the spring of 2020 (Table 5). When estimated using the Robust Design models (under the original AICc analysis) and corrected for the unmarked portion of the population, N_total varied from approximately 94 (SE = 12.68; 95% CI 72-122) to 204 individuals per primary period (SE = 22.85; 95% CI 164-54) (Table 5; Figure 4). Seasonal abundance estimates exhibited a recurring oscillatory pattern across years, generally increasing from winter into spring or summer before declining again in fall and the following winter (Figure 4). Seasonal variability in abundance across years was greatest in summer (SD = 26.5), intermediate in winter (SD = 20.1), and substantially lower in spring (SD = 9.2) and fall (SD = 8.2), indicating that temporal fluctuations in abundance were not evenly distributed across seasons (Table 6). Annual means (Figure 5) increased from 123 in 2013 to a peak of 172 in 2020 and subsequently stabilized within a narrower range of 150–166. Mean abundance across all 10 years was 156.23 (SE = 23.38; 95% CI 117-209). Model-averaged marked-abundance estimates were sensitive to the change in supported capture-probability structure when ĉ > 1. Across the 40 primary periods, estimates under ĉ = 4 differed from the original estimates in both directions, with a mean absolute difference of 16.9 marked dolphins, equivalent to 15.4% of the original estimates (Supplementary Table S3).

Table 5. Summary of abundance estimates based on model selection using AIC_c.

n (all): total number of unique dolphins marked per primary period; n (distinct): number of unique, marked non-calf dolphins captured in the core study area per primary period (i.e., those included in analyses) Nhat_best: best abundance estimates based on the best supported single model using robust design; N_hat: model-averaged abundance across the top 3 models; N_total: total abundance after a delta correction for the unmarked proportion of the population.

Table 6. Mean seasonal abundance estimates and corresponding detection probabilities.

Values are based on the AIC_c model-averaged, delta corrected N_total abundance estimates.

Under the original AICc analysis, mean detection probability (p) for presumed residents was 0.36 compared to 0.17 for non-residents. Temporal variation in detection varied widely between and within primary periods (Supplementary Table S4), but was strongly synchronized across mixture groups (Figure 6), indicating that variation is driven by sampling period-level factors rather than group-specific processes. Estimates from presumed residents ranged from 0.19 (0.07 SE; 95% CI 0.05-0.33) in the summer of 2015 (primary period 11) to 0.59 (SE = 0.12; 95% CI 0.35-0.83) in the winter of 2013 (primary period 1). The median and mean capture probability across all model-averaged primary periods for presumed residents were both 0.36 (SE = 0.10; 95% CI 0.16-0.55). Detection probability for presumed residents exhibited modest seasonal variation, with slightly higher mean values in spring (mean p = 0.39) and winter (mean p = 0.38), intermediate values in summer (mean p = 0.35), and the lowest values in fall (mean p = 0.32) (Table 6). However, substantial overlap in confidence intervals among seasons indicated that seasonal effects (averaged across years) on detectability were relatively weak compared to variation among primary periods (i.e., within-year seasonal variation). Consistent with the shift toward simpler capture probability structures, when recalculated under ĉ = 4, model-averaged detection probabilities exhibited little variation among primary periods, averaging 0.469 (range: 0.465–0.472) for presumed residents and 0.179 (range: 0.177–0.181) for presumed non-residents (Supplementary Table S5). Thus, increasing assumed overdispersion primarily reduced support for temporal variation in detection, while the distinction between the two mixture groups remained.

Because survival structures differed across the three highest-ranked models, survival estimates were not model-averaged and are therefore reported separately (Supplementary Table S6). Under the original AICc analysis, the best fit model included a survival term that varied by season (S.season). Apparent survival was highest during winter and fall (winter: 1.0 ± 5.62 × 10^-8, 95% CI 0.99-1.0; fall: 1.0 ± 6.54 × 10^-8, 95% CI 0.99-1.0) and lowest during summer (0.86 ± 0.04; 95% CI 0.75-0.92). This pattern is consistent with stranding records indicating higher mortality during summer months than during fall or winter (Wells, 2010). The mean apparent survival across all seasons was 0.95 (mean 95% CI 0.88-0.98) (Figure 7). The second-best fit model varied by the transience effect only (S.A). As expected, estimated apparent survival for the group modeled over the first interval after initial detection was low at 0.68 (SE = 0.08; 95% CI 0.51–0.82), consistent with what would be expected for animals with single sightings or that visited the area infrequently (presumed non-residents), whereas apparent survival over subsequent intervals was 0.96 (SE = 0.01; 95% CI 0.94–0.98), consistent with presumed residents. In the third-ranked model (S.Ayear), apparent survival for presumed residents varied from 0.88 to 1.00 among years and averaged 0.96, with the lowest estimates occurring in 2021 (0.88) and 2020 (0.90). Under extreme assumed overdispersion (ĉ = 4), the three highest-ranked survival structures were S(A), S(.), and S(season). Point estimates (Supplementary Table S6*) for structures shared with the original analysis changed little: S(A) estimates were 0.96 for subsequent intervals (presumed residents) and 0.67 for the first interval (presumed non-residents). Seasonal apparent survival (S.season) estimates followed the same pattern as the original seasonal model—highest in winter and fall (1.0) and lowest in summer (0.86). Corresponding point estimates differed from the original estimates by ≤0.01, although standard errors for estimable parameters increased by approximately 92–99%. The new constant-survival model estimated apparent survival at 0.96 (SE = 0.02; 95% CI 0.89–0.98). Overall, despite differences in how survival was structured among the top models and shifts in model ranking under assumed overdispersion, apparent survival estimates for presumed residents remained consistently high, whereas estimates for the first interval after detection were substantially lower. Thus, inference regarding overall resident survival was robust, although support for the specific source of variation in survival differed among models.

4 Discussion

This study provides the first multi-year set of abundance estimates for Sarasota and Little Sarasota Bays using a robust design capture–mark–recapture framework. Model selection (interpreted based on AICc analyses) favored those incorporating seasonal variation in apparent survival; strong non-random availability; transience effects on temporary emigration; and strong heterogeneity in detection probabilities, explained by variation among primary periods and between mixture groups (i.e., presumed residents and non-residents). Results adequately reflect what we would expect for this system (Fazioli et al., 2006; Wells et al., 2025; Wells and McHugh, 2025): the presence of a relatively stable resident core community, with seasonal movement driving fluctuations in both apparent survival and abundance for the resident community.

After the first year of the study, there was little variability in mean abundance estimates (ranging from 123 to 172 over the 10 years), suggesting a relatively consistent population size over the ten-year study period. Confidence intervals around annual abundance estimates were relatively wide (~90–100 individuals) but exhibited little variability in range from year to year; Figure 5, Table 5). The wide intervals can be explained by the combined effects of detection heterogeneity, temporary emigration, and aggregation across primary periods. Specifically, heterogeneous detection probabilities increase uncertainty in scaling from marked to total population size, while non-random availability means that abundance is inferred from a subset of individuals present within the study area at any given time. Additionally, annual estimates combine variability and uncertainty from multiple primary periods, resulting in compounded uncertainty rather than averaging effects.

While annual abundance estimates were relatively consistent over time, there were strong within-year oscillations (ΔN_total = ± 30–54 individuals between consecutive primary periods), which followed a repeating seasonal pattern across years (Figure 4). Temporal variation in abundance was not distributed evenly across seasons—variability in abundance was greatest in summer and lowest in fall (Table 6). Detection probability also exhibited seasonal structure, with slightly higher mean detection probabilities in spring and winter (p = 0.39 and p = 0.38, respectively) and lower values in fall (p = 0.32); however, seasonal differences in detection were relatively small overall (~0.06 difference between highest and lowest seasonal means) and did not mirror the pronounced cyclical pattern observed in abundance estimates. In addition, mean variability in abundance estimates across years was lowest in fall despite it being the season with the lowest mean detection probability, suggesting that seasonal changes in detectability were not the primary driver of temporal variation in abundance. Instead, periods of low detection probability (Supplementary Table S4) were associated primarily with reduced precision in abundance estimates—reflected by wider confidence intervals during periods such as summer 2015 (primary period 11 in Figure 4). Together, these results suggest that detection heterogeneity influenced the magnitude and uncertainty of abundance estimates, whereas the temporal structure in abundance itself was driven primarily by seasonal changes in availability. This interpretation is further supported by the strong non-random temporary emigration estimates and asymmetric transitions between availability states for presumed residents versus non-residents. Seasonal redistribution of animals within and outside the core study area likely contributed more to fluctuations in abundance than seasonal variation in detectability alone. Of note, the original AICc analysis supported substantial variation in detection among primary periods, whereas analyses assuming higher levels of overdispersion favored substantially simpler detection structures. Consequently, both the degree of temporal heterogeneity in detection and the resulting primary period abundance estimates were sensitive to the assumed variance-inflation factor. In contrast, the distinction between mixture groups and model-averaged temporary emigration point estimates remained comparatively stable. Because the evaluated values of ĉ were assumed rather than empirically estimated, the sensitivity analysis does not identify which detection structure is most biologically accurate, but it demonstrates that inference regarding temporal variation in detection and abundance depends partly on the assumed degree of overdispersion.

Tyson Moore and Wells (2016) estimated abundance for 2015 using the same core study area, same photo-screening process, and a similar robust design modeling framework, although their best-supported model did not include temporary emigration or detection heterogeneity. Their annual mean abundance estimate across primary periods (157.5 dolphins) was very close to our 2015 annual mean estimate (153.2 dolphins), supporting broad agreement in estimated population size. Both analyses also showed an increase in summer followed by a decline in fall, which is in alignment with the seasonal redistribution patterns first described by Irvine et al. (1981). The primary differences involved the structure of the best-supported models and the degree of uncertainty associated with the summer estimate. With ten years of data and the inclusion of heterogeneity and movement structure, our models were able to estimate temporary emigration patterns that were likely difficult to resolve from only four primary periods sampled within a single year. As such, the no-emigration model result reported by Tyson Moore and Wells (2016) likely reflected limited power to estimate movement processes rather than true absence of movement. Failure to account for temporary emigration can bias apparent survival downward when temporarily unavailable individuals are treated as permanent losses from the sampled population. This was evident–Tyson Moore and Wells (2016) reported a mean apparent survival estimate of 0.87 which was substantially lower than in the present study (mean = 0.95) and lower than long-term survival estimates previously reported for Sarasota Bay (mean apparent survival = 0.94–0.95 from 1995–2020; Schwacke et al., 2023). Using the long-term dataset likely also improved the stability of parameter estimation during periods of low detection probability (by incorporating substantially more encounter histories). Generally, as detection probability (p) decreases, abundance estimates and associated uncertainty are inflated because a larger proportion of the population is inferred to be unseen. That is evident when comparing our results from summer 2015 to that reported by Tyson Moore and Wells (2016), a season with particularly low detection probabilities (p = 0.19 and p = 0.09, respectively). We report a notably lower abundance estimate (164 vs. 218, respectively) and less inflated confidence intervals (95% CI: 105.6–253.4 vs. 128.7–369.8, respectively). Despite these differences, both studies produced broadly similar annual abundance estimates and seasonal patterns.

Abundance estimates from this study were broadly consistent with the minimum abundance estimates reported by (Wells et al., 2025) using census data, which documented annual counts ranging from 108 to 184 individuals, with recent estimates approaching ~200 as of 2019. Although the two approaches estimate different quantities—minimum known residents versus total abundance corrected for unmarked individuals—the overlap in magnitude is substantial. The upper range of N_total in this study (~172 - 220) closely aligns with the upper bounds of census estimates. This agreement confirms that robust design capture–recapture methods can be used to produce comparable estimates, when heterogeneity (including transience) is accounted for. Bounded abundance estimates are also consistent with projections from Lacy et al. (2021), which modeled the Sarasota population under a carrying capacity (K) of approximately 200 individuals and predicted stabilization near this level (mean ~197 individuals at equilibrium). The clustering of annual means just under 200 (~150–172 with the exception of 2013, which was 123.3) combined with the absence of an increasing trend across ten years (or notable annual variation of any kind after the second year; Figure 5) are consistent with a population operating near carrying capacity. However, results here should be interpreted with caution given that similar constraints on abundance could arise for other reasons (e.g., environmental variability).

Although the uncertainty in abundance estimates is consistent with expectations for an open population, the high uncertainty evident in our annual estimates limits the sensitivity of this modeling approach to detect subtle temporal trends. The differences in estimates among years (150–170 individuals) were small relative to the width of confidence intervals (117–204 on average). This limitation is particularly relevant when evaluating responses to environmental perturbations such as Karenia brevis red tide blooms. For example, there were severe red tide events in 2005–2006 and 2018-2019 (Berens McCabe et al., 2021; Gannon et al., 2009). Lacy et al. (2021) reported increased calf and juvenile (up to five years old) mortality and population declines associated with both events (in 2006 and 2019). The severe red tide from fall 2018 through winter 2019 coincides with a slight decrease in the seasonal abundance estimates reported here (Fall 2018 = 166; Winter 2019 = 148; ΔN_total = −17), but this decrease is small compared to changes between the same two seasons either before or after the event (2017-2018 ΔN_total = -29.67; 2020-2021 ΔN_total = +0.31) and less than the range of typical fluctuations observed across the entire time series (ΔN_total = ±30–50). Previous work in Sarasota Bay demonstrates that ecological responses to harmful algal blooms can be delayed, with prey assemblages recovering over 1–3 years following bloom events (Berens McCabe et al., 2021; Gannon et al., 2009). Similarly, dolphin demographic responses may follow perturbations rather than occur immediately (Wells et al., 2025). However, abundance based on our results increased slightly from 2019 (N_total = 159.8) to 2020 (N_total = 172.2) and remains similar for the next two years (2021 N_total = 161.1; 2022 N_total = 165.7). Due to data limitations related to parameter identifiability, the models we were able to explore may have lacked the ability to detect population-level effects of episodic disturbances such as red tide.

Capture probabilities (mean = 0.36, 95% CI: 0.19-0.59) were either comparable to (e.g., mean = 0.31 in Charleston (Speakman et al., 2010) and up to ~0.6 in Indian River Lagoon; Schwacke et al., 2023) or in some cases lower than those reported from other studies in the southeast (e.g., St. Joseph Bay, Florida: mean = 0.29-0.69; the Southern Georgia Estuarine system: mean = 0.01-0.87; Schwacke et al., 2023). Our capture probabilities were lower than would be expected for this system given the high level of survey effort, the relatively few non-resident animals that utilize the inshore system seasonally (Fazioli et al., 2006), how well known the population is, and the large number of sighting records available for many individuals of the local community (e.g., Wells et al., 2025). For example, in an unpublished study¹ that tracked 183 calves born between 2004 and 2019 from time of birth until 2020 (or until death), the average time interval between consecutive sightings was less than 65 days for 94% (172) of individuals and less than 30 days for 84% (153) of individuals.

The relatively modest capture probabilities reported here may, in part, reflect our program's rigorous photo quality standards. Although most bottlenose dolphin capture-mark-recapture studies in the southeast use the same photo-screening processes, our program's criteria for scoring photo quality are notably strict. For example, photos of fins will get marked as “partial” fins in our FinBase catalog if the base of the dorsal fin is not completely visible in a given image, even if the leading-edge dips slightly below the waterline but remains visible. Such images are excluded during PQ screening if no other image from the sighting contains the complete dorsal fin. Some research programs are more lenient on this criterion, allowing minor fin obstructions before downgrading to “partial” (see Urian et al., 2015 for a report on the variation in PQ scoring among researchers and resulting biases). We may therefore exclude a larger proportion of marked fins during photo-quality screening than programs applying less stringent criteria. The extent of data loss during screening demonstrates the potential effect of these standards on detection histories. 18.4% of the original 7,956 unique sighting records from 493 individuals were removed during screening, resulting in 250 individuals in the analytical dataset. Notably, the effect of photo-quality screening alone was substantial. Prior to screening for photo quality or single sightings, 318 unique, marked non-calves were detected in the core study area, totaling 5,587 unique sightings and 3,160 primary period detections. Screening these sightings for photo quality alone reduced the number of unique animals to 290, resulting in the removal of 38.8% (2,167) of unique sightings and 26% (823) of primary-period detections. At least one primary period detection was lost for 66.6% (193) of the 290 individuals, with a maximum of 15 (out of 31) lost for a single individual.

Temporary unavailability of animals within the surveyed area likely further contributed to the relatively modest capture probabilities. The study area represents only the core portion of the Sarasota Bay system. The local dolphin community is known to utilize the adjacent coastal habitats and is therefore temporarily unavailable for detection within the study area during these periods (Fazioli et al., 2006; Wells and McHugh, 2025). Many of the animals deemed “non-resident” were those sighted at the northern end of our survey range (north of Anna Maria Island), an area known to overlap with adjacent communities from South Tampa Bay (Urian et al., 2009), Manatee River, and/or Terra Ceia Bay (Wells et al., 1987; Wells, 2014). Given the slightly overlapping dolphin community ranges, we are unlikely to ever fully account for detection heterogeneity using a Robust Design model. These factors collectively contribute to uncertainty and limit the ability to detect subtle changes in population size using this modeling approach.

Future work should compare multiple modeling approaches to evaluate tradeoffs between maintaining comparability with other inshore bottlenose dolphin studies and improving inference regarding seasonal movement, temporary emigration, and true demographic change. Assigning individual covariates of “resident” and “non-resident” would better control for heterogeneity in capture probabilities and thereby decrease model uncertainty but would reduce comparability given that long-term sighting histories from which to make those distinctions (at the individual level) are not available for most systems. Alternatively, expanding analyses to include all monthly survey data instead of just seasonal data and incorporating coastal sightings would also likely increase detection probabilities for the resident dolphins when they are using nearshore coastal habitats; however, adding coastal sightings will also increase the number of non-residents included in the dataset, which would again increase heterogeneity (and decrease detection probabilities). One way to address the issue of whether or not to include coastal sightings is to assign location (inshore vs. coast) as state and model changes in state directly using multistate robust design models (e.g., Chabanne et al., 2017). This approach would allow explicit estimation of movement between habitats and improve understanding of metapopulation structure overall. Finally, spatially explicit capture-recapture models have been increasingly used to more effectively handle heterogeneity in capture probabilities in the estimation of density and abundance of dolphins (e.g., Balmer et al., 2018; McDonald et al., 2017; Pirotta et al., 2014). Spatially explicit approaches combine capture histories with spatial locations to estimate capture probabilities based on the distance from identified activity centers for each individual. As such, this approach offers information on how individuals use space, which can be used to model transitions across habitat types and predict responses to spatial and temporal disruptions. Such approaches would provide a more powerful framework for assessing spatial dynamics and responses to localized disturbances, which are critical for management in coastal systems.

In summary, this study demonstrates that abundance estimates for bottlenose dolphins in Sarasota Bay are relatively stable in magnitude but exhibit seasonal oscillations driven by availability and movement dynamics. Although wide confidence intervals limit the ability to detect trends or disturbance effects, the consistency of estimates with long-term census data and theoretical expectations provides confidence in the overall magnitude of population size. These findings underscore the importance of accounting for movement and heterogeneity in capture–recapture analyses and highlight both the strengths and limitations of robust design approaches for understanding population dynamics of resident dolphin communities.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars.2026.1914202/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Lorenzo Mari, Polytechnic University of Milan, Italy

Reviewed by: Raluca Ioana Bancila, Romanian Academy, Romania

Shannon Barber-Meyer, Cascadia Research Collective, United States

Footnotes

Allen et al. (2021). "Ch-ch-changes: Dorsal fin mark acquisition rates in a community of bottlenose dolphins, Tursiops truncatus, in Sarasota Bay, Florida, USA." 24th Biennial Conference on the Biology of Marine Mammals, Palm Beach, Florida, August 1-5, 2022.