Journal of Herpetology, Vol. 51, No. 1, 102–108, 2017 Copyright 2017 Society for the Study of Amphibians and Reptiles
Estimating Occurrence and Detection Probabilities for Stream-Breeding Salamanders in the Gulf Coastal Plain JENNIFER Y. LAMB,1,2 J. HARDIN WADDLE,3
AND
CARL P. QUALLS1
1
Department of Biological Sciences, The University of Southern Mississippi, 118 College Drive #5018, Hattiesburg, Mississippi 39406, USA. 3 U.S. Geological Survey, Wetland and Aquatic Research Center, 700 Cajundome Blvd., Lafayette, Louisiana 70506, USA.
ABSTRACT.—Large gaps exist in our knowledge of the ecology of stream-breeding plethodontid salamanders in the Gulf Coastal Plain. Data describing where these salamanders are likely to occur along environmental gradients, as well as their likelihood of detection, are important for the prevention and management of amphibian declines. We used presence/absence data from leaf litter bag surveys and a hierarchical Bayesian multispecies single-season occupancy model to estimate the occurrence of five species of plethodontids across reaches in headwater streams in the Gulf Coastal Plain. Average detection probabilities were high (range = 0.432–0.942) and unaffected by sampling covariates specific to the use of litter bags (i.e., bag submergence, sampling season, in-stream cover). Estimates of occurrence probabilities differed substantially between species (range = 0.092–0.703) and were influenced by the size of the upstream drainage area and by the maximum proportion of the reach that dried. The effects of these two factors were not equivalent across species. Our results demonstrate that hierarchical multispecies models successfully estimate occurrence parameters for both rare and common streambreeding plethodontids. The resulting models clarify how species are distributed within stream networks, and they provide baseline values that will be useful in evaluating the conservation statuses of plethodontid species within lotic systems in the Gulf Coastal Plain.
Lungless salamanders (Family Plethodontidae) constitute a substantial proportion of the vertebrate biomass (Burton and Likens, 1975; Semlitsch et al., 2014) and play important roles in energy cycling (Best and Welsh, 2014) in many temperate ecosystems. Plethodontids are the focus of numerous studies, but the majority of these have involved populations in the Northwest (Sheriden and Olson, 2003; Olson and Weaver, 2007) or in the Appalachians rather than in the Gulf Coastal Plain (Means, 2000). The Gulf Coastal Plain is a physiographic province with a unique history, topography, and suite of climates and habitats (Kirkman et al., 2007). The environmental gradients that shape species occurrence, or the importance of any particular gradient, may differ among these provinces. In light of ongoing amphibian declines (Stuart et al., 2004), including the enigmatic decline of some species within the Gulf Coastal Plain (Means and Travis, 2007; Maerz et al., 2015), herpetologists must collect and analyze baseline data describing where species are likely to occur and at what frequency we might expect to encounter populations within an area. These efforts will help characterize long-term trends and test explicit hypotheses, enabling us to detect, monitor, and possibly prevent declines in the future. Three environmental gradients, stream size (Waldron et al., 2003), topography (Marshall and Camp, 2006), and stream impermanence (Means, 2000), are cited as important factors affecting the occurrence of plethodontids in the Southeastern United States. More specifically, each is suspected to shape assemblages within the Gulf Coastal Plain (Means, 2000), but hypotheses regarding these associations have not been empirically tested. Although sometimes correlated, these environmental gradients can vary independently of one another, and the impact of each on occupancy should be considered separately. A variety of ecological factors that could affect the occupancy of stream-breeding plethodontids change along the stream size gradient, such as the composition of the fish community and water temperature (Vannote et al., 1980). Plethodontids can 2
Corresponding author. E-mail:
[email protected]
DOI: 10.1670/16-050
persist at sites containing predatory fishes (e.g., Lepomis [Petranka, 1983]), but species may mitigate predation pressure by occupying smaller streams in which predator gape-size is limited (Vannote et al., 1980) and in which overall densities may be restricted by lower or intermittent flows (Richardson and Daney, 2007). Small streams also tend to have cooler temperatures during the summer months, and recent work in midAtlantic drainages indicates that the probability of occurrence for some plethodontids increases with decreasing average water temperatures (Grant et al., 2014). These and other factors that vary along the stream size gradient may act in a complex, synergistic fashion to shape species occurrence in the Gulf Coastal Plain. The landscape through which a stream flows may determine how species are organized longitudinally within the stream. The Coastal Plain lacks extreme relief, but it does contain relatively steep hills, bluffs, and deep ravines (Kirkman et al., 2007) in which conditions can substantially differ from those in flat bottomland habitats (e.g., temperature, rate of flow). Species of plethodontids within the Gulf Coastal Plain may be relegated to specific habitats along this gradient if they are physiologically constrained by recent evolutionary history (e.g., thermal tolerances) (Bernardo and Spotila, 2006; Kozak and Wiens, 2012). Competitive exclusion also may play a role in the distribution of species along the topographic gradient, either in the arrangement of species when moving perpendicularly away from the stream (e.g., Hairston, 1949) or in their distribution between steep, headwater origins and swampy downstream habitats (e.g., Means, 1975). The relationship between occupancy and stream hydroperiod has been investigated for salamanders in the Northwest (Olson and Weaver, 2007), but relatively few studies in the Southeastern United States ask how stream impermanence affects plethodontids. Species in the Appalachians can be absent from or occur at lower abundances in reaches that are experimentally dried (Currender et al., 2014), but all plethodontids are not necessarily affected by drying in the same way. Ephemeral or semipermanent streams, which may be intermittent along their length, are common in the Gulf Coastal Plain and can contain
SALAMANDER OCCUPANCY IN THE GULF COASTAL PLAIN Dwarf (Eurycea quadridigitata), Three-lined (Eurycea guttolineata), and Southern Dusky Salamanders (Desmognathus auriculatus) (Petranka, 1998; Bruce, 2005; Lamb, pers. obs.); however, we do not understand whether occupancy changes for these species along this drying gradient. Larval periods among biphasic species in the Gulf Coastal Plain range from 4 mo to more than 2 yr, and considerable intraspecific variation exists in this trait (Dundee and Rossman, 1989; Petranka, 1998; Bruce, 2005). Occupancy of ephemeral sites by other plethodontids may be precluded or limited by metamorphic parameters (e.g., developmental rate) for species derived from lineages that morerecently occurred in stable stream habitats (e.g., Desmognathus) (Bruce, 2005). Adult salamanders that survive periods of low water levels can buffer a population from extirpation, but these populations can persist for only a limited period of time in the absence of recruitment (Price et al., 2012). Here, we use Bayesian methods and a hierarchical multispecies single-season model (Ke´ry and Royle, 2008; Royle and Dorazio, 2008; Waddle et al., 2013) to estimate the effects of stream size, topography, and stream impermanence on salamander occupancy across reaches (n = 60 sites) in small to medium headwater streams in the Gulf Coastal Plain. Hierarchical occupancy models quantify relationships between the occurrence of a species and environmental covariates while also incorporating detection probabilities (MacKenzie et al., 2002, 2006; Royle and Dorazio, 2008). These models, which have been applied to both anurans (Walls et al., 2011; Waddle et al., 2013; Lehtinen and Witter, 2014) and caudates (Bailey et al., 2004; Nichols et al., 2008; Grant et al., 2009; Walls et al., 2013), are powerful tools that allow us to account for imperfect detection, a persistent problem for many species of amphibians (MacKenzie et al., 2002) because of the influence of sampling conditions (Walls et al., 2011, Waddle et al., 2013) or study design (Walls et al., 2013; Grant et al., 2014; Lehtinen and Witter, 2014). MATERIALS
AND
METHODS
Study Area and Site Selection.—We surveyed 60 sites across 32 streams in the Pascagoula River Drainage in South Mississippi, USA (Fig. 1). We defined a site as a 50-m long reach and selected sites along two habitat gradients, stream size and surrounding topography (i.e., ravine vs. bottomland), in an effort to represent the habitat diversity across reaches within small to medium 1stand 2nd-order streams in the Pascagoula. Twenty-eight of the 32 streams contained two sites each. We separated sites within the same stream by at least 100 m of stream length. Stream-breeding plethodontids do not generally make long-distance movements (Cecala et al., 2009; Wells, 2010), and individuals likely did not move between reaches during this study. Species are not necessarily continuously distributed along a stream, even across short distances (e.g., Currender et al., 2014), and studies similar to ours have separated sites by comparable intervals (Nichols et al., 2008). Habitat variables of interest can differ across these distances in streams, though not necessarily to the same extreme as they might between streams. Eleven sites occurred in the Bienville National Forest, 47 in the De Soto National Forest, and two in the Ward Bayou Wildlife Management Area. Data Collection.—We visited sites from May through July of 2012 or 2013. Each site was sampled three times (sampling occasions), and sites within the same stream were sampled within the same summer. Sites contained five leaf litter bags (litter bags) (Pauley and Little, 1998; Waldron et al., 2003) separated by 10 m. We constructed litter bags from a double-
103
FIG. 1. Streams (dots) in the Pascagoula River Drainage containing sampled reaches (n = 60). The inset depicts part of the Gulf Coastal Plain with the study area enclosed in a thick black box. State names are abbreviated and National Forests are shaded grey.
layered 70 · 70 cm square of plastic wildlife netting with pores 1.5 cm in diameter (Waldron et al., 2003), filled them with leaf litter in situ, and tied each bag to the bank. To sample a litter bag, we quickly lifted it from the water while sweeping a dip net beneath it and then placed the litter bag in a large plastic container (Waldron et al., 2003; Mattfeldt and Grant, 2007). After checking the net for salamanders, we poured water over the bag in the container and agitated the bag for 60 sec. We then poured the contents of the container through the net. We agitated bags as described until no salamanders were dislodged after two consecutive attempts. We identified all salamanders to species and released them close to their point of capture, except for a small number of individuals that we collected for use in other studies. At each site, we collected habitat data describing the three factors of interest: stream size, surrounding topography, and stream impermanence. We recorded stream size using two metrics, wetted-width (Waldron et al., 2003) and upstream drainage area (ha) (Snodgrass et al., 2007). We measured wetted-width to the nearest 1 cm at distances of 5, 15, 25, 35, and 45 m within each site during each sampling occasion. The average of these data for each site constitutes the width covariate (Width). We used the U.S. Geological Survey National Hydrography Dataset and ArcGIS 10.3 (ESRI, Redlands, California, USA) to estimate upstream drainage area (DArea). Sites within the same stream had the same value for DArea because they were too close for this metric to appreciably differ. We used the measurement tool in ArcScene 10.0 and 1/3 arcsecond (~10 m) digital elevation models, provided by the U.S.
104
J. Y. LAMB ET AL.
TABLE 1. A priori hypotheses regarding the probability of occurrence (W) and environmental gradients. Negative signs (-) mark relationships for which the probability of occurrence (W) is predicted to decrease as the value of the covariate increases. Positive signs (+) mark relationships for which W is predicted to increase as the value of the covariate increases. Zeros indicate that there is no predicted association between this species and covariate. Species
Desmognathus conanti Eurycea cirrigera Eurycea guttolineata Eurycea quadridigitata Pseudotriton ruber vioscai
Stream size Stream impermanence Topography
0 0 -
0 0 -
+ 0 0 0 +
Geological Survey National Elevation Database, to estimate the upstream drainage area for streams that were not available in the National Hydrography Dataset. We measured the percent slope of the streamside habitat along a 10-m line perpendicular to the reach on each side of the bank at meters 0, 25, and 50 within each site with a clinometer. These data were then averaged for each site (Topography). Sites in this study dried either completely, partially, or never during the summer in which they were surveyed. We estimated stream impermanence for each reach by quantifying the maximum proportion that dried during the sampling season. We took three equally spaced depth measurements across the wetted-width at meters 5, 15, 25, 35, and 45 within a site, giving a total of 15 depth measurements per site for each of the three sampling occasions. We calculated the proportion of points equaling zero during each sampling occasion and used the maximum of these three values to describe stream impermanence (Dry). We used three sampling covariates that we hypothesized could influence detection probability as a consequence of our choice of survey method (i.e., litter bags) including litter bag submergence, sampling date, and the type and proportion of instream cover present within each site. Waldron et al. (2003) noted that the number of metamorphosed salamanders caught in litter bags is negatively correlated with the proportion of the bag that is submerged, suggesting that transformed salamanders may not utilize the entirety of the stream channel. This possibility, combined with our sampling during the summer months in which multiple species undergo metamorphosis and exit the stream (Lamb, pers. obs.), could result in lower detection probabilities for species with shorter larval periods. With this in mind, we estimated litter bag submergence for each bag to the nearest 25% prior to checking the bag for salamanders and then averaged these percentages for each sampling occasion and site (Submerge). To control for the effect of the time of year, we also included the number of days since 1 May as a detection covariate (Day). Waldron et al. (2003) suggest that the availability of natural cover within the stream is negatively correlated with the likelihood that salamanders would utilize litter bags (i.e., lower densities in bags because of greater availability of suitable refugia elsewhere). Anecdotal evidence suggests, however, that streams containing more in-stream cover generally support greater densities of plethodontid larvae (Lamb, pers. obs.), increasing the detectability of this life stage. We quantified the amount and type of in-stream cover available to salamanders using five equally spaced, 4-m-wide belt transects crossing each reach. Within these transects we visually assessed, to the nearest 1, 5, or 10%, the area covered by bare substrate (i.e., sand, silt,
TABLE 2. Number of individual plethodontid salamanders caught in leaf litter bags across 60 sites. Transformed individuals include those completing metamorphosis (i.e., gill and tail-fin absorption nearly complete). Species
Larvae
Transformed
Total
Desmognathus conanti Eurycea cirrigera Eurycea guttolineata Eurycea quadridigitata Pseudotriton ruber vioscai Total
0 1,637 67 85 17 1,806
17 105 100 35 2 169
17 1,742 167 120 19 2,065
and/or clay bottom), leaf-litter, woody debris, aquatic vegetation, and roots. Totals for a transect could sum to >100% because in-stream cover can describe three-dimensional structure. We calculated the average proportions of each type of instream cover from belt transects for each site. We then used these data in a principal components analysis of covariates and the resulting site scores along the first principal axis in the model (Cover). Data Analysis.—We pooled sites sampled in different years into a single dataset. We compared the amount of in-stream habitat variation (i.e., between sites within the same stream) to that observed between streams by performing principal components analyses and calculating the average pairwise Euclidian distances between points. We used Width, DArea, Topography, and Dry data, and the average proportions of each type of in-stream cover in these analyses. Data for sites within the same stream were averaged when comparing streams, and we excluded streams containing only one site (n = 4). Our model estimates probability of occurrence for five species of stream-breeding plethodontid salamanders and is similar to the multispecies model in Waddle et al. (2013). In this model, the detection history (1 = present, 0 = not detected) for each species (i = 1, 2, . . .5) across multiple sampling occasions (j = 1, 2, 3) at each sampled site (k = 1, 2, . . .60) is summarized as yijk and is contingent upon occurrence (z) and detection probability (p) so that yijk ~ Bern (pijk * zik). There are two possible interpretations when a species is not detected; either it does not occur at that site (zik = 0), and therefore was not available for detection, or it occurs there (zik = 1) but researchers failed to observe it (MacKenzie et al., 2006). Occurrence is a latent variable distributed Bernoulli on the probability of occurrence (W), zik ~ Bern (Wik). Site and sampling covariates are used to separately model W and p, respectively, through application of the logit link function and the effect size for site (b), and sampling (a) parameters are estimated for each species (Royle and Dorazio, 2008). We used four covariates to model the probability of occurrence: Width, Topography, Dry, and DArea; logit (Wik) = b0 + (b1 * Width) + (b2 * Topography) + (b3 * Dry) + (b4 * DArea). We centered and scaled data for each covariate. Table 1 lists our a priori hypotheses for how each of the five species might respond to each of the four site covariates. These hypotheses were based on relationships described in the literature (Petranka, 1998; Means, 2000; Waldron et al., 2003) and on personal observations of stream-breeding plethodontids in the Gulf Coastal Plain. We included three covariates to model detection probability: Submerge, Day, and Cover; logit (pijk) = a0 + (a1 * Submerge) + (a2 * Day) + (a3 * Cover). We centered and scaled Submerge and Day and completed data transformations and principal components analyses in R (R Core Team, 2014).
SALAMANDER OCCUPANCY IN THE GULF COASTAL PLAIN TABLE 3. Estimated mean detection probabilities and 95% BCI for stream-breeding plethodontids. Species
Desmognathus conanti Eurycea cirrigera Eurycea guttolineata Eurycea quadridigitata Pseudotriton ruber vioscai
p (SD)
0.461 0.942 0.459 0.432 0.634
(0.168) (0.026) (0.060) (0.125) (0.136)
Lower 95% BCI Upper 95% BCI
0.134 0.883 0.341 0.191 0.347
0.764 0.984 0.577 0.666 0.868
Multispecies occupancy models are community-level models in which ecologically similar species are treated as ‘‘random effects,’’ meaning that species-specific effects for a parameter are drawn from a common distribution. This approach allows for heterogeneity in detection and occupancy among species and can result in more-precise estimates, particularly when dealing with rarer species (Ke´ry and Royle, 2008; Walls et al., 2011; Waddle et al., 2013). We used Bayesian analysis with uninformative priors to estimate model parameters. We applied uniform distributions from 0 to 1 to priors for community-level occurrence and detection probabilities and normal distributions with a mean of 0 and variance equaling 5 to priors for the effect(s) of site covariates and 8 for sampling covariates, respectively. We incorporated Kuo and Mallick’s (1998) inclusion parameter to identify which covariates improved the fit of the model. Values for inclusion parameters were binomially distributed on 0.5. We fit this multispecies model using the Markov chain Monte Carlo method in WinBUGS (ver. 1.4.3) (Spiegelhalter et al., 2003) and called WinBUGS from R using the package R2WinBUGS (Sturtz et al., 2005). We used three parallel chains 200,000 iterations in length with a burn-in length of 100,000 and a thinning rate of 10. We assessed Markov chains using the Gelman-Rubin statistic (R-hat) and number of effective sample sizes (n.eff) in WinBUGS and by visually examining trace plots. We generally considered R-hat 100 to be indicative of acceptable convergence and mixing (Gelman and Hill, 2007). Posterior distributions were not affected when we tested more-diffuse priors. We report the mean values and 95% Bayesian credible intervals (BCI) of the posterior distributions for those covariates that were maintained in the final model after Kuo and Mallick (1998) selection. We report descriptive data as means 6 SD. RESULTS We captured 2,065 plethodontid salamanders in litter bags (Table 2). We detected Southern Two-lined Salamanders (Eurycea cirrigera) in 60%, E. guttolineata in 30%, and E. quadridigitata in 13% of the total of 180 sampling occasions across all 60 sites. We detected Spotted Dusky Salamanders (Desmognathus conanti) and the Southern Red Salamander (Pseudotriton ruber vioscai) in approximately 6% of the total
105
number of sampling occasions. The only non-plethodontid salamander detected was the Western Lesser Siren (Siren intermedia nettingi) (n = 2). Our study sites had relatively small wetted-widths (186.1 6 90.3 cm) and upstream drainage areas (528 6 460 ha). Drainage areas for 11 sites (7 streams) were estimated using digital elevation models in ArcScene. Topographies of many sites were flat or gently sloping, but some sites included steeper terrain (6.10 6 11.91% slope). The majority of sites contained water throughout the summer or only 10% of the reach dried (0.13 6 0.29 maximum proportion dry). In-stream habitat variation (3.45 6 2.68 pairwise Euclidian distance) was less than that observed between streams (14.88 6 8.75 pairwise Euclidian distance), but sites within the same stream were not identical. For example, sites within the same stream did not dry to equivalent proportions (n = 16 sites). Four sites dried completely during the second sampling occasion, two of which remained dry for the remainder of the study. These sites have detection histories with missing response data (i.e., NA values) which are estimated by WinBUGS (Ke´ry 2010). Estimated mean detection probabilities (p) ranged from 0.432 6 0.125 to 0.942 6 0.026 (Table 3) and were lowest for E. quadridigitata. The 95% Bayesian credible intervals (95% BCI) were widest for D. conanti and P. ruber vioscai (Table 3). The mean posterior values for the Kuo and Mallick (1998) inclusion parameters of the sampling covariates Submerge, Day, and Cover were low at 0.006, 0.027, and 0.684, respectively, and their 95% BCIs substantially overlapped zero, suggesting that these covariates did not account for any appreciable variation in detection probabilities. Minimum occupancy, defined as the proportion of sampled sites at which the species was detected at least once, ranged from 0.08 to 0.65 (Table 4). Estimates of the mean finite probability of occurrence (W) across sampled sites ranged from 0.092 6 0.017 to 0.703 6 0.060, and the 95% BCI was greatest for E. quadridigitata (Table 4). Two site covariates, DArea and Dry (Table 5), were retained in the model with mean posterior values for their inclusion parameters of 1.000 and 0.996, respectfully. These values were
