Mobile acoustic transects miss rare bat species: implications of survey method and spatio-temporal sampling for monitoring bats

Method 1: mobile transect surveys

We used the only existing paved park road from the main ENP entrance gate to Flamingo (64 km), which traverses the three vegetation communities: pinelands, prairies and mangroves (Fig. 1A). Transects were driven at a constant 32 km/h, beginning 30 min after sunset (Loeb et al., 2015; United States Fish and Wildlife Service, 2012), for a survey duration of two hours. We surveyed the transect nine times in the dry season and ten times in the wet season, each within a 40-night sampling period (ca. once every five nights), alternating the start location between each end of the transect road for each survey. We selected a sample effort much greater than the two survey nights per maternity season recommended by the widely-implemented NABat protocol (Loeb et al., 2015), to estimate the number of survey nights necessary to detect maximum species richness in the study area. We did not survey on nights when average wind speeds exceeded 24 km/h (United States Fish and Wildlife Service, 2012), but allowed for wind gusts up to 30 km/h due to consistently high coastal wind conditions. Due to frequent but often short-duration rain events in south Florida, we paused each survey when precipitation began and waited for it to pass. If the survey was paused for more than 15 mins cumulatively, we ended the survey and repeated it the following night. We assigned each bat detected during a mobile transect to the vegetation community in which it was recorded, using that file’s associated GPS coordinates (ArcMap v10.2.2; Florida Cooperative Land Cover Map v3.1).

Method 2: single strategic stationary surveys

We established one easily accessible, strategically-located stationary monitoring station designed to remain permanently in each of the three vegetation communities: pinelands, prairies, and mangroves (Fig. 1A). To maximize bat detections and the quality of recordings, we located each detector within 10 m of a water source, but away from vegetative clutter and flat surfaces prone to acoustic echoes (Britzke, Gillam & Murray, 2013; Loeb et al., 2015). Detectors were powered by solar panels, mounted with custom made structures constructed from PVC pipes and angle iron, and programmed to record from 30 mins before sunset to 30 mins after sunrise (12-hr mean survey duration/night over the study). We set the detectors to record for 40 nights in each season.

Method 3: spatially replicated stationary surveys

To assess how increasing sample effort spatially would affect estimated bat species richness and how that may differ between structurally distinct vegetation communities, we focused on pinelands (small, dry, closed canopy system) and prairies (vast, wet, open canopy system), two relatively easy to access vegetation communities. We did not sample in mangroves due to high water levels year round that would require additional equipment (e.g., airboats) for these surveys to be conducted. We sampled a total of six locations in pinelands (2–12 km apart) and six locations in prairies (8–45 km apart), distributing sample locations as evenly as possible throughout each vegetation community (≥ 300 m from the habitat edge), while ensuring that each location was accessible by foot year round (10–300 m from a secondary road or path, 100–3,500 m from a primary road) (Fig. 1A). We mounted detectors and microphones on painter’s poles (3 m above ground) and selected sites where the recording space above microphones was unobstructed by vegetative clutter. We programmed detectors to record from 30 mins before sunset to 30 mins after sunrise (12-hr mean survey duration/night over the study). In each season, we conducted two 20-night sampling sessions. During each 20-night sampling session we placed three detectors in pinelands and three detectors in prairies and then moved these detectors to six new locations for the second 20 night sampling session. The same 12 locations were surveyed during both the wet and dry seasons.

Survey method

First, we did a simple comparison of the species (and total species richness) detected by all three methods in each survey location in each season. To account for the limited temporal sampling by mobile transects relative to stationary surveys, and to identify potential reasons for differences in species detections, we subsequently restricted this comparison to: (a) the nights, (b) the nights and times when mobile transects were conducted.

Next, we compared nightly species richness between mobile transects and single strategic stationary surveys only, to test the effectiveness of these two methods across the three primary vegetation communities. To account for the greater number of nights sampled by the stationary surveys, we limited our comparison to only the nights when mobile transect surveys were conducted. Although the number of recording hours per night was less for mobile transects than stationary surveys, we used ‘survey night’ as our sample unit based on existing survey protocols and what is typically used in the field (Loeb et al., 2015; United States Fish and Wildlife Service, 2012). In the practical application of these survey methods, mobile transects are never conducted all night, and stationary surveys are generally always conducted for the full night. We constructed generalized linear mixed-effects models (GLMMs), with vegetation community and season as random effects (function glmer, R package lme4 Bates, Maechler & Bolker, 2012). We modeled all GLMMs with a Poisson distribution for count data, after assessing model fit using standard graphical diagnostics of residuals (Zuur et al., 2009) and testing for over-dispersion (function dispersion_glmer, R package blmeco (Korner-Nievergelt et al., 2015). To test for the effect of survey method, we compared two nested models (one with survey method included and one without) using a likelihood ratio test (function anova, R Stats package).

Sample effort

To evaluate how spatially increasing sample effort within vegetation communities affected the estimated species richness, we compared the two stationary survey methods in pinelands and prairies and used the entire sampling period (40 nights in each season). We compared the nightly species richness that was detected by single strategic and spatially replicated stationary surveys. We constructed GLMMs (Poisson distribution), with survey method and vegetation community as fixed effects and season and individual detector as random effects (categorical variables), and used likelihood ratio tests to evaluate the effect of survey method. We then used the same approach to test for differences between survey methods within pinelands and prairies to determine if there was a difference in the effectiveness of spatial replication in two structurally distinct vegetation communities.

Species accumulation curves

To further compare survey methods and to estimate the minimum temporal and spatial sampling effort necessary to detect maximum expected species richness, we generated sample-based rarefaction species accumulation curves (SAC) (Gotelli & Colwell, 2001), using the function ‘specaccum’ in the R package Vegan. We used the default ‘exact’ method, which estimates the expected species richness using a sample-based rarefaction method (Chiarucci et al., 2008; Kindt, Vandamme & Simons, 2006) and is often known as the Mao Tau estimate (Colwell et al., 2012). Maximum expected species richness is estimated at the point that SACs reach an asymptote, or the probability of detecting new species approaches zero. Rarefaction curves can be used to compare the efficacy of different sampling methods within an area(Ellison et al., 2007; Longino, Coddington & Colwell, 2002). We used a simple, but approximate, method to compare SACs among survey methods by looking at the degree of overlap in confidence intervals (CI) (Gotelli & Colwell, 2011). To do this, we calculated unconditional standard deviations (Colwell, Mao & Chang, 2004; Colwell et al., 2012), which are not conditioned on the empirical data set but are instead based on an estimation of the extrapolated number of species in the regional species pool (i.e., assumes the sample is randomly drawn from a larger, unknown, species assemblage (gamma diversity); function specpool). We then plotted 84% confidence intervals (ci = 1 in function specaccum) for each SAC, which are more appropriate for tests of overlap than traditional 95% CI, assuming approximately equal CIs (Gotelli & Colwell, 2011; Payton, Greenstone & Schenker, 2003). To compare SAC overlap between single strategic stationary surveys and mobile transects, we limited the sample effort to the same number of nights for each method by generating a random subsample, without replacement, of single strategic stationary survey nights (nine in the dry season and 10 in the wet season). To determine the temporal sample effort required to detect all bat species using each survey method in each of the two seasons, we used the full sample period and plotted SACs with number of survey nights as the sample unit. Multiple detectors within each survey method (single strategic = 3; spatially replicated = 12) were pooled for these analyses. To assess how required temporal sample effort varied by vegetation community, we plotted SACs within pinelands and prairies for the two stationary survey methods. Finally, to determine the minimum spatial sample effort required to detect all bat species using the spatially replicated stationary surveys in each season, we plotted SACs in pinelands and prairies with number of detector sites as the sample unit.

Spatial—survey locations

The same total bat species richness (N = 9) was detected with both stationary survey methods (Table 2). Spatially replicated stationary surveys detected significantly lower mean nightly bat species richness (2.74 ± 0.06) than did single strategic stationary surveys overall (3.60 ± 0.10; χ²(1) = 8.16, p = 0.004). This result was also found within pinelands (3.23 ± 0.09 vs. 4.71 ± 0.15; χ²(1) = 12.18, p < 0.001), but not within prairies (2.24 ± 0.06 vs. 2.60 ± 0.09; χ²(1) = 0.94, p = 0.33). In both seasons, the overlapping CIs of the SACs indicated that there was no significant difference in expected species richness or rate of species accumulation between single strategic stationary surveys and spatially replicated stationary surveys overall (Figs. 3A–3B) and within pinelands (Figs. 3C–3D). In contrast, within prairies in the dry season, the SACs of spatially replicated stationary surveys had higher expected species richness and steeper curves than the SACs from the single strategic stationary surveys (Figs. 3E–3F). In the wet season, overlapping CIs indicated no difference in expected species richness between single strategic stationary surveys and spatially replicated stationary surveys, but spatially replicated stationary surveys did not reach an asymptote within the 40-day sampling period indicating insufficient temporal sample effort to detect all species. When evaluating the optimal number of detector sites in spatially replicated stationary surveys, asymptotes were reached on the SACs with only four detector sites in pinelands during both the dry and wet season (Fig. 4). In contrast, spatially replicated stationary surveys in prairies did not reach asymptotes with the six detector sites in either season, indicating that spatial sampling effort was not sufficient.

Temporal—survey nights

The minimum number of sample nights required to detect expected species richness varied among survey methods, seasons and vegetation communities (Fig. 3). Maximum expected species richness (N = 9 species) was reached only during the wet season with stationary surveys in both pinelands and prairies combined (Fig. 3B). The minimum sample effort required to achieve this was 38 nights for both single strategic stationary surveys and for spatially replicated stationary surveys. Although SAC asymptotes were also reached within the dry season and within each vegetation community, fewer than nine species were detected in each of these categories alone. An asymptote was not reached in prairies for spatially replicated stationary surveys in the wet season after 40 nights of sample effort (Fig. 3F).

Spatial—survey locations

Contrary to what we expected, spatially replicated stationary surveys detected the same total species richness but lower mean species richness per night than single strategic stationary surveys, with no difference in the rate of species accumulation or estimated species richness overall. Various studies have shown that multiple detectors are necessary to adequately represent the diversity of bats across and within vegetation communities (Ciechanowski et al., 2007; Duchamp et al., 2006; Froidevaux et al., 2014; Jung et al., 2012; Skalak, Sherwin & Brigham, 2012). Our findings suggest that the more labor-intensive survey method of sampling spatially within vegetation communities was not necessary, and that a single detector, strategically located in pinelands and prairies to maximize bat detections, was sufficient. Skalak, Sherwin & Brigham (2012) suggested that detecting all bat species using a single detector was possible, but improbable in their study area (Mojave Desert, Nevada). Our placement of the single strategic detectors adjacent to bodies of water was likely responsible for the effective sampling of bat species richness. Although water is abundant in ENP, much of it is brackish and/or may be inaccessible to bats for drinking due to dense vegetation impeding flight paths. Because most bats require open sources of fresh water in order to swoop down to drink (Kunz, 1982; Taylor & Tuttle, 2007), bats in ENP may congregate around the limited open bodies of fresh water. Our findings support existing recommendations to select stationary survey sites that maximize potential bat activity, which are often near water (Loeb et al., 2015). Many ecosystems worldwide are either truly water limited for bats (e.g., xeric habitats, Korine et al., 2016) or may appear water rich yet be limited in sources of water that are accessible to bats (e.g., due to dense vegetative clutter, Ciechanowski et al., 2007; Jackrel & Matlack, 2010). In cases where there are no water bodies available for detector placement or there is limited knowledge of where bats are most active, spatial replication may be able to substitute for strategic placement of detectors.

Increased spatial sampling provided no additional benefits to estimates of species richness in pinelands but it increased the rate at which new species were detected and the maximum estimated species richness detected in prairies. Because prairies are open and wet (Lodge, 2010), a given open water source in prairies may be relatively less attractive to bats than an open water source in the drier, structurally complex pinelands. Thus, sampling near water in only one location in the prairies may not be sufficient to fully capture species richness in this community. Similarly, when evaluating the optimal number of detector locations and distribution needed to spatially represent bat communities in ENP, we again observed differences between the two vegetation communities sampled. Our finding that maximum species richness was reached (SAC asymptote) in pinelands with four sites (for 20 nights each) indicates that we adequately sampled spatial heterogeneity in this vegetation community in both dry and wet seasons. Although we found that the single strategic detector was sufficient to detect maximum species richness in pinelands, our estimates of minimum spatial sample effort for pinelands are applicable to monitoring schemes in other similar woodland communities that may not contain (or have access to) sources of water at which to locate detectors. In contrast, spatially replicated stationary surveys in prairies did not ever reach an asymptote in either season, indicating that more than six detectors (for 20 nights each) would be needed to fully represent the bat species richness in this vegetation community. Skalak, Sherwin & Brigham (2012) similarly found that at least six detector locations were necessary (for ca. 30 nights each) to detect 90% of bat species in a wetland system. We suggest that prairies in ENP required a greater sample effort than pinelands due to their much larger area and consequently higher heterogeneity in microhabitat and patterns of bat activity and species distributions, similar to findings by Moreno & Halffter (2000) of a greater sample effort required for bats in a large heterogeneous landscape than in small homogeneous habitats.

Temporal–survey nights

We found that all nine bat species could be detected with two strategic stationary detectors (one in each of two vegetation communities) after 38 nights of sampling in the wet season when species richness was highest. In contrast, Whitby et al. (2014) detected all 12 bat species in their study area in Illinois after only 12 nights of sampling (three sample events of four nights each) with stationary detectors. A complete inventory of bat species can require a substantial sample effort (Moreno & Halffter, 2000). Skalak, Sherwin & Brigham (2012) recommended 20–30 sampling nights to detect 80–90% of the estimated bat species richness (12 species) in Nevada. Froidevaux et al. (2014) recommended a range of 12 to 33 nights for 90% of estimated species richness (16 species) in Switzerland. In our study, we detected 90% of the estimated species richness (eight species), after only seven nights of sampling with two strategically placed detectors in the wet season. The relatively low sample effort required to detect 90% of species in ENP, yet high sample effort required to detect all species is likely due to the presence of rare bat species, which have lower detection probabilities and higher variation in activity levels (Skalak, Sherwin & Brigham, 2012). The variation in optimal survey duration by these studies illustrates that blanket recommendations for monitoring protocols may not be appropriate for all regions, and may need to be fine-tuned based on site-specific factors and project objectives.