Predation risk and prey responses exhibit fluctuations in space and time. Seasonal ecological disturbances can alter landscape structure and permeability to influence predator activity and efficacy, creating predictable patterns of risk for prey (seasonal risk landscapes). This may create corresponding seasonal shifts in antipredator behavior, mediated by species ecology and trade-offs between risk and resources. Yet, how human recreation interacts with seasonal risk landscapes and antipredator behavior remains understudied.

Goals – In South Florida, we investigated the impact of a seasonal ecological disturbance, specifically flooding, which is inversely related to human activity, on interactions between Florida panthers (Puma concolor coryi) and white-tailed deer (Odocoileus virginianus). We hypothesized that human activity and ecological disturbances would interact with panther-deer ecology, resulting in the emergence of two distinct seasonal landscapes of predation risk and the corresponding antipredator response. We conducted camera trap surveys across southwestern Florida to collect detection data on humans, panthers, and deer. We analyzed the influence of human site use and flooding on deer and panther detection probability, co-occurrence, and diel activity during the flooded and dry seasons. Flooding led to decreased panther detections and increased deer detections, resulting in reduced deer-panther co-occurrence during the flooded season. Panthers exhibited increased nocturnality and reduced diel activity overlap with deer in areas with higher human activity. Supporting our hypothesis, panthers' avoidance of human recreation and flooding created distinct risk schedules for deer, driving their antipredator behavior. Deer utilized flooded areas to spatially offset predation risk during the flooded season while increasing diurnal activity in response to human recreation during the dry season.

Significance to the field – We highlight the importance of understanding how competing risks and ecological disturbances influence predator and prey behavior, leading to the generation of seasonal risk landscapes and antipredator responses. We emphasize the role of cyclical ecological disturbances in shaping dynamic predator-prey interactions. Furthermore, we highlight how human recreation may function as a 'temporal human shield,' altering seasonal risk landscapes and antipredator responses to reduce encounter rates between predators and prey. 

Study Site

We conducted our study in the Big Cypress Swamp physiographic region of southwestern Florida, encompassing the Florida Panther National Wildlife Refuge (FPNWR) and Big Cypress National Preserve, including Bear Island (BI) and the Addition Lands (AL; Fig 1, Supporting Information 1). This region served as the primary restoration zone for Florida panthers (US Fish and Wildlife 2008). It experienced distinct flooded and dry seasons (Fig 1), along with seasonal storms and hurricanes (Duever et al. 1994; Abernathy et al. 2019). The regional topography exhibited minimal relief, characterized by slight ridges separating flat basins interspersed with depressions that could retain standing water throughout the dry season.

Public access varied across FPNWR, BI, and AL, with hiking trails being present in all areas and experiencing human site use. BI allowed off-road vehicle (ORV) access and hunting, while AL permitted public recreational access and limited hunting but prohibited ORV use. FPNWR, on the other hand, did not allow public access or hunting but allowed ORV use for management purposes. Importantly, while human access levels varied across our study area, representing a gradient of human disturbance, the overall level of human presence was relatively low compared to other studies examining the influence of humans on wildlife (Riley et al. 2003; Gehrt, Anchor & White 2009). In recent years, concerns over excessive deer harvest have led to regulatory changes aimed at limiting access, reducing the total harvest of adult males, and eliminating the harvest of spotted fawns and females (Schortemeyer et al. 1991; Florida Wildlife Code, 68A-13.004). Adult males that do not meet antler criteria, which require at least one antler to be five inches or longer and have at least two points, are not allowed to be hunted. All hunting activities occur during the flooded season (Florida Wildlife Code, 68A-13.004).

While infrared-flash camera traps are commonly used in wildlife research, recent studies have suggested that white-flash cameras may not significantly affect detection probabilities for many species. Further, white-flash cameras may even provide advantages such as color nighttime photos that improve species identification, an important aspect in determining species' diel activity patterns (Herrera et al. 2021; Ladd, Meek & Leung 2022). Therefore, we deployed 180 motion-triggered white-flash trail cameras (HCO Outdoor Products, model SG565FV, Norcross, GA, USA) in FPNWR, BI, and AL, creating three camera grids within our study area. Each camera grid covered an area of more than 29 km², and the grids were spaced more than 13 km apart. Within each grid, we positioned 40 trail cameras on trails, approximately 700 m apart, and an additional 20 camera traps around 250 m from each other, located off the trails (Fig 1, Supporting Information 1). We monitored the cameras from January 1, 2015, to December 31, 2017, following the maintenance and data retrieval protocols outlined in Crawford et al. (2019).

Spatial co-occurrence and detection probability

We aimed to investigate panther-deer co-occurrence and its relationship with human recreational use and flooding across hydrological seasons. Co-occurrence captures the shared site use by both predators and prey, providing insights into their spatial interactions. To achieve this, we utilized the hierarchical co-occurrence model developed by Rota et al. (2016).

The Rota et al. (2016) model extends MacKenzie's et al. (2002) single-species model by linking a detection model with a partially observed latent process co-occurrence model. This hierarchical model comprises three sub-models: species-specific conditional detection probability (p_sit), marginal occupancy (ψ_i), and species co-occurrence (ψ₁₁). To estimate detection probabilities (p_sit), we required repeated sequential visits to the same camera site (MacKenzie et al. 2017). Since our camera data was continuous, we summed species- and sex-specific camera detections across seven continuous days to represent distinct sampling occasions. For each occasion, we categorized each camera location as occupied (1) or not occupied (0) for male and female deer and panthers. We developed competing models incorporating anthropogenic and environmental variables that we believed influenced the detection probabilities, marginal occupancy, and co-occurrence of male and female deer and panthers (Table 1).

To assess the impact of recreational activity on detection probability, we summed weekly, effort-corrected human detections per camera trap night as a measure of human activity rate. Flooding effects on detection were evaluated using a weekly mean estimate of a ‘surface water index’ as represented by spatially-explicit, daily water levels previously developed for our study area (Abernathy et al. 2021). Additionally, since panthers are more likely to be detected on trails (Crawford et al. 2019), we included the on-or-off trail status as a variable in the species and sex-specific detection models.

To examine the influence of flooding on deer-panther co-occurrence, we averaged surface water levels across each hydrological season for each camera-year. Next, we averaged effort-corrected human detections per trap night across each hydrological season for each camera-year to account for species responses to human use. Since our primary focus was deer-panther co-occurrence, the sex and species-specific occupancy (ψ_i) models included the year of our study to account for potential differences in species occupancy or site use across years. Additionally, to address the lack of comparability across camera locations (Supporting Information 2), we treated each unique camera-year combination as a separate 'camera' in our analysis (Fuller, Linden & Royle 2016).

It is worth noting that occupancy models assume geographically closed populations during sampling periods, but this assumption can be relaxed in cases where population changes occur due to random chance or when ψ is interpreted as the probability of animal use within the sampling area (MacKenzie et al. 2017). In our study, we selected a sampling area based on deer home range size, which is relatively small compared to the panther's home range (US Fish and Wildlife 2008). Thus, marginal occupancy (ψ_i) and co-occurrence (ψ₁₁) in our models represent individual and shared site use rather than true occupancy in the formal sense.

In our modeling framework, detection probabilities were used to estimate site co-occurrence (shared site use), considering imperfect detection. Therefore, we initially developed competing models to explain species-specific detection while keeping occupancy constant (intercept-only models). We compared these models using Akaike's information criterion (AIC) to determine the best variables explaining variation in detection probabilities (Burnham & Anderson 2002). The top detection model was then used to develop competing models to explain female and male deer co-occurrence with panthers (Table 1). Similarly, to determine which variables best-explained variation in species shared site use, we compared competing models using AIC (Burnham & Anderson 2002). All explanatory variables used in detection and co-occurrence modeling were scaled and centered, and we observed no strong correlations among the variables (Table 1, Supporting Information 1). We performed data curation and analyses in R version 4.2.2 and implemented our detection and co-occurrence modeling using the 'unmarked' package (version 1.2.5; Fiske & Chandler 2011).

Temporal activity 

We aimed to explore the temporal dynamics of panther-deer interactions in response to human activity and seasonal flooding. To investigate this, we employed several analytical approaches. First, we sought to quantify the effects of hydrological season and human use on male and female deer and panther temporal activity. To accomplish this we first classified each sex and species-specific camera detection into diurnal (one hour past sunrise to one hour before sunset), crepuscular (sunrise to one hour after sunrise and one hour before sunset to sunset), or nocturnal (sunset to sunrise) categories using the maptools package (version 1.1-5; Bivand & Lewin-Koh 2018). Next, we calculated the upper (≥ 66% distribution), median (< 66% and > 34% distribution), and lower (≤ 33% distribution) quantiles for human and vehicle detections for each camera grid (Hyndman & Fan 1996). We classified cameras within respective upper, median, and lower quantiles as high, median, and low human use, respectively, across all hydrological season-years. To ensure comparability, we excluded off-trail cameras when estimating human site use due to very low human detections in those areas.

Next, we calculated the proportion of detections for each deer sex and species across diurnal, crepuscular, and nocturnal time periods during each hydrological season-year. We constructed a candidate set of beta regression models (Ferrari & Cribari-Neto 2004) to investigate the influence of hydrological seasons (flooded and dry) and human use (high, median, and low) on diel activity for each sex and species (Table 2, Supporting Information 1) utilizing the betareg package (version 3.1-4; Cribari-Neto & Zeileis 2010). We compared competing models using AIC corrected for small sample size (Burnham & Anderson 2002). To summarize the effects of hydrological seasons and human use factors on the proportions of detections across times of day (diurnal, crepuscular, nocturnal), we estimated least-squares means for the factor combinations within each sex and species top model using the emmeans package (version 1.8.2-090003; Lenth et al. 2021). Finally, we calculated 95% confidence intervals for estimated least-squares means using the Satterthwaite method.

Additionally, we sought to understand how hydrological seasons and human recreational use influenced shared temporal activity between deer of both sexes and panthers. To achieve this, first we estimated diel activity overlap between panthers and deer at each camera location by fitting kernel density functions to timestamps of deer and panther detections. We then estimated the coefficient of overlap, a quantitative measure ranging from 0 (no temporal overlap) to 1 (identical temporal activity patterns; Ridout & Linkie 2009), using the overlap package (version 0.3.4; Meredith, Ridout & Meredith 2018). We applied a smoothing parameter of one if we had 50 or fewer detections and Δ hat of four if we had more than 50 detections to reduce bias in coefficient and standard error estimates (Ridout & Linkie 2009). Next, we quantified human recreational activity at each camera by calculating annual mean human detections per total camera trap nights, while mean annual surface water at each camera location represented flooding. We constructed linear models investigating sex- and species-specific diel overlap using every combination of mean human activity and mean flooding serving as explanatory variables (Table 3, Supporting Information 1). We used AIC corrected for small sample size to determine our top model (Burnham & Anderson 2002). We scaled and centered mean human activity and flooding values, and ensured they were not strongly correlated (r = -0.03).

To address small sample size concerns at certain camera locations, we calculated diel activity overlap between male deer-panthers and female deer-panthers across all cameras within each level of human recreational use (high, medium, and low) for both hydrological seasons (flooded and dry). Additionally, considering the influence of life history phenology in deer activity and the potential impact of hunting on antipredator responses (Main, Weckerly & Bleich 1996; Ruckstuhl & Neuhaus 2000; Cromsigt et al. 2013), we estimated the coefficient of overlap for each sex-species combination across all cameras within each level of human recreational activity (high, medium, and low) for each deer biological season (fawn birthing [Jan – Mar]; fawning rearing/antler growth [Apr – Jun], rut [Jul – Aug], gestation/post-rut [Sep – Dec]). Confidence intervals for each coefficient of overlap estimate were generated using a smoothed bootstrap approach with 10,000 iterations in the overlap package (version 0.3.4; Meredith, Ridout & Meredith 2018). We performed all analyses in R version 4.2.2 (R Core Team 2022).

R version 4.2.2