Due to access limitations and variation of suitable landscape types within study sites, we did not select our camera trap locations randomly. We typically placed camera traps along levees and roads bordering habitats and movement corridors suitable for Tegus as informed by resource use documented by Klug et al. (2015), where the authors found that the vegetated levees within aquatic habitats (i.e., marshes) facilitated Tegu activity and movement. This meant that many cameras were spatially aligned in a relatively linear fashion. We also placed camera traps in residential and agricultural areas to assess potential variation of Tegu occupancy with regards to land use. The average straight-line distance between a camera trap and the nearest camera trap was 954 meters (range: 24–6014 m; SD ± 983 m). There were cases when cameras placed in proximity (i.e., <30 m) were positioned on opposite sides of canals at minimal straight-line distances. However, the actual distance an animal would have to travel between camera traps was likely much greater, as canals may serve as a semi-permeable barrier to Tegu movement (Klug et al. 2015). Our camera trap placement also facilitated trap checking and maintenance, as levees provided a means of access within sites despite water inundation during the wet season, while deep-water canals often otherwise necessitate specialized equipment for access. Additionally, concurrent research examining spatial ecology of Tegus through radio telemetry suggested that Tegus avoid inundated marsh (Mason 2022).

We collected SD cards from our unbaited camera traps approximately once per month. Upon retrieval, we examined camera traps to ensure they were in working order, removed vegetation or other obstructions of the camera view, if present, and replaced camera trap batteries as needed. For baited camera traps, we visited camera traps and collected SD cards approximately once every 2 weeks. Upon these visits, we repeated the same procedure, but also replaced lures (i.e., chicken eggs) as needed. Following SD card collection, a member of our team manually checked camera trap photos to detect whether a Tegu was present in the photo (1 = yes, 0 = no). For data quality and assurance, we also had a different person “proof” or visually inspect each photo a second time.

Occupancy Analysis

Data preparation. In many cases, our cameras were active year-round. However, we excluded photos taken on camera traps between 1 November and 15 February of each year. Tegus generally brumate during cooler months, starting as early as September and ending in February or early March in southern Florida (Currylow et al. 2021, McEachern et al. 2015), so we selected these sampling periods to encompass the entirety of the Tegu active season in our study site. This is further supported by our own camera trap data, as we observed detections decrease in frequency during September and October, with the latest detection occurring on 7 November. Our earlier detections in the calendar year began to increase in late February, with our earliest detection observed on 20 February. During brumation, we suspect that the probability of detection is near 0 if not 0, even though occupancy of our selected sites may not change.

We generated binary detection histories in R (v.4.3.1) (R Core Team 2021) using the package camtrapR (v 2.0.3) (Niedballa et al. 2016). We delineated our camera trap effort into sampling occasions consisting of 14 consecutive days, as we found that Tegus had low daily detection probabilities and this length of time allowed us to reduce non-detections and optimize modelling (Sollmann 2018). We included all camera trapping effort, meaning that if a camera was operating at any point during the 14-day period, we treated this period as a sampling occasion (see Supplemental Figure 1 in Supplemental File X, available online at http://www.eaglehill.us/SENAonline/suppl-files/nXX-X-XXX-Xxxxxx-sX.)). 

Covariate Data Collection

Site covariates. For each camera trap, we measured habitat type and distance (m) to landscape features (Table 1) using shapefile layers informed by the Florida Cooperative Landcover Map (CLC). We calculated the Euclidean distance (m) from each camera trap to the boundary of polygons representing urbanization, agricultural lands, and to 2 areas we deemed relevant to the invasion history of Tegus in South Florida. One of these areas is a suspected introduction site for Tegus in South Florida (hereafter referred to as “source population”). The source population site is characterized by a quarry bordered by an agricultural matrix and is north of many of the removal efforts targeting Tegus in South Florida. The second area or landmark of interest (hereafter referred to as the “UF core trapping area”) included a parcel of land where several live traps were being operated by University of Florida to capture Tegus. This area was located at an underpass of US Highway 1 and is where linear habitat features (i.e., Highway US1 and vegetated shrubland along a road) intersect within a largely inhospitable landscape (i.e., aquatic). The UF core trapping area is located at the northeastern boundary of an area deemed the “core management area” by Udell et al. (2022), where live trapping conducted by various agencies (i.e., University of Florida, United States Geological Survey, and the National Park Service) has proved productive for Tegu removal, and we assumed that relative abundance in this area was high. The core management area is made up of several levee systems in the Southern Glades and is approximately 4 km south of the source population. We suspected that the linear habitat features (i.e., levees or roads) connecting the core management area and the source population could aid in the dispersal of Tegus. The UF core trapping area, however, may serve as one of few dispersal pathways that would allow for Tegus to cross Highway US 1. Additionally, the intersection of linear habitat features may act as a large-scale drift fence, funneling Tegu movement within the Southern Glades to this area, potentially increasing occupancy probability.  

We classified habitat into 3 categories: levee, terrestrial, and disturbed. We classified levee sites as cameras that were on a levee that was bordered by aquatic habitats (i.e., marsh or canals) on both sides. For cameras placed on levees that had aquatic features on one side, but terrestrial landscape features (i.e., forested area, naturally, or semi-naturally occurring vegetative cover) on the other side that would aid Tegu dispersal or movement, we classified this land use type as “terrestrial”. Classification of disturbed habitat type included sites where camera traps were placed in, or directly adjacent to, anthropogenic features (i.e., public roads, houses, and agricultural fields) that may hinder Tegu movement. We grouped camera trap sites into these 3 categories by visually inspecting satellite imagery to identify habitat characteristics and dominant landscape features of the areas adjacent (i.e., within a buffer of ~50 m) to the camera sites.

Detection covariates. We suspected that whether the camera trap was baited, temperature, rainfall, the number of days a camera trap was active during the 14-day sampling occasion, and the ordinal date would influence detection probability of Tegus. As ectotherms, Tegus are highly susceptible to environmental conditions, which likely influence their activity patterns and detectability. We sourced temperature and rainfall data from the Florida Automated Weather Network station located in Homestead (25˚30’45”N, 80˚30’11”W). For each sampling occasion, we aggregated mean daily temperature and rainfall data to coincide with our 14-day sampling occasions and calculated the mean for each sampling occasion. We used the mid-point of each 14-day sampling occasion to assign an ordinal date to each sampling occasion.

Yearly site covariates. Many multi-season or multi-year occupancy studies track occupancy probabilities and detection for conservation-based information (e.g., using occupancy as a proxy to track variation in population size and distribution of vulnerable species) (De Wan et al. 2009, Farris et al. 2017, Gojiman et al. 2015). In these cases, using year or other variables that may change between primary sampling seasons (e.g., canopy cover or forest cover) is often appropriate. However, working with invasive species often requires their detection and subsequent removal, so including information about removal efforts may be better applied to understanding how occupancy changes over time in response to management. For this reason, we used data from live traps that were operated by collaborating agencies at the same time as our camera traps to calculate seasonal covariates. We calculated the Euclidean distance (m) to the nearest known live trap for each camera trap in each sampling season and the capture per unit effort (CPUE) of the nearest live trap to investigate changes in occupancy and detection probabilities over time as a response to removal efforts. We had knowledge of and capture data for a total of 1,129 livetraps in our study area, with an average of 282 traps operating per year (range = 249–314; see Supplemental Figure 2 in Supplemental File X, available online at http://www.eaglehill.us/SENAonline/suppl-files/nXX-X-XXX-Xxxxxx-sX.). Live traps were distributed by collaborating agencies along levees, berms, and other areas suitable for Tegus with many of them aggregated in the core management area (n = 690), with 18 of these traps located at or within a 100 m radius of the UF core trapping area. We calculated live trap CPUE by dividing the number of Tegus captured in a trap by the total number of trap nights for that trap.

Modeling Framework

Occupancy models are applied to estimate occurrence (i.e., occupancy) of species at a site while taking into account imperfect detection (Kéry et al. 2010). In a traditional occupancy model, sites must be sampled on multiple occasions because non-detections do not necessarily mean that a site is not occupied (MacKenzie et al. 2002). Occupancy models can incorporate variables that may influence detection (detection covariates) and change between sampling occasions or between sites (site covariates) to account for heterogeneity (MacKenzie et al. 2003). Dynamic occupancy models attempt to answer similar questions (i.e., what proportion of sites are occupied, and what is the probability of detecting a species if it is present?), but over time. In addition to detection covariates and site covariates, a dynamic or multi-season occupancy model may include covariates that are not expected to change in between sampling occasions but may change between sampling seasons to account for local colonization (i.e., a site becoming occupied that was not in previous seasons) or local extinction (i.e., a site that was once occupied was not found to be occupied in following seasons).

We created dynamic occupancy models using the R package unmarked (v.1.2.5) (Fiske and Chandler 2011) to see how occupancy of Argentine Black and White Tegus changed over the course of 4 sampling seasons (2016–2020). Prior to model creation, we standardized our detection covariates (temperature, rainfall, ordinal day, camera trapping effort per 14-day sampling occasion), site covariates (distances to landscape features and areas of interest, specifically urbanization, agriculture, the UF core trapping area and the source population), and colonization and extinction covariates (distance to nearest live trap and the CPUE of that trap). We created models applying a stepwise procedure (e.g., Stewart et al. 2022) which entails creating models at several steps where covariates are either added to or removed from the model based on their performance (i.e., AIC_c score) as demonstrated by Button et al. (2019). At step 1, we created 9 models, including a null model and a model for each detection covariate including camera trap bait status, temperature, precipitation, camera trapping effort for each 14-day sampling occasion, and ordinal day with a quadratic effect to determine which covariate best explained detection probability. At this step, we also created a model for each of our seasonal covariates (distance to nearest live trap and CPUE of nearest trap), as these could be considered factors that may also influence detection. For example, the presence of a live trap with bait could act as a lure and thus increase the probability of detection. Additionally, the presence of a very productive live trap may decrease detection locally as individuals are removed from the landscape. We used the R package AICcmodavg (v.2.3-1) (Mazerolle 2023) to select top performing models using Akaike’s Information Criterion corrected for small sample sizes (AIC_c), and we considered models with ∆ AIC_c  ≤ 2 to be top performing models (Burnham and Anderson 2002). This step allowed us to identify 1 or more covariates that best explain detection to be included in models in step 2.

Once we identified which model best predicted the probability of detection (p), we moved on to step 2 and built 6 models that included covariates retained in the top performing detection model. This included a model for each of our site covariates (habitat type, distance to UF core trapping area, distance to source population, distance to urbanization, and distance to agriculture) as well as one model where occupancy was assumed to be constant across sites and seasons (i.e., intercept only).

In the third and final step, we included covariates included in the top performing detection model (step 1) as well as those included in the top performing occupancy model (step 2).  At this step we included our seasonal covariates (CPUE and distance to nearest live traps) to identify which seasonal covariate best explains colonization and extinction of Tegus at our sites. These 7 models included various combinations of our seasonal covariates, as well as 1 model where colonization and extinction were assumed to be constant across sites and sampling seasons (i.e., intercept only) (Table 2). We calculated annual estimates of site occupancy using the “smoothed” function in unmarked to use the finite sample estimate of site occupancy.