An acoustic-based method for locating maternity colonies of rare woodland bats

Phase 1: Acoustic methodology development

We collected acoustic data between May and September 2019 in 13 ancient semi-natural and replanted woodlands in the counties of West Sussex, Wiltshire, and Herefordshire in the UK. We based site selection on (i) previous records of known barbastelle colonies (n = 6) or (ii) the presence of ancient woodland where colonies were believed to be absent (n = 7). Sites were categorised as having barbastelle colonies or not on the basis of historical evidence from radio-tracking studies, trapping and/or bat box checks completed by third parties. We considered historical evidence as any surveys conducted within the five years preceding our study period (2014–2019). In addition, we analysed the temporal activity patterns in the acoustic data that was subsequently collected: we expected sites with colonies present to show clear peaks in activity close to the emergence time of barbastelles, as well as near dawn. In cases where we were unable to discern whether colonies were present or absent with a reasonable level of confidence, we categorised the sites as ‘unknown’ and used them to test our models. The woodlands selected ranged in size from 14 to 122 ha, with elevations between 15 m (50°57′02″N, 0°36′26″W—Burton & Chingford Ponds Nature Reserve) and 203 m (52°4′43″N, −2°58′8″W—Moccas Park National Nature Reserve). Climatic conditions in the region are mild and wet with mean summer (May–September) temperatures between 14.4 and 15.2 °C and mean precipitation between 275 and 284 mm (Hollis et al., 2019).

We collected acoustic data at 146 survey points using full-spectrum passive Song Meter SM2BAT static detectors with SMX-US or SMX-U1 omni-directional microphones (Wildlife Acoustics, Maynard, MA, USA; see Article S1 for full detector settings). We spaced the detectors approximately 150 m apart along trails and rides, features known to be favoured by barbastelles (Greenaway, 2004), deploying 5–15 detectors per woodland depending on size and accessibility of the site. A strict grid was not used because some areas were inaccessible; and others suggest that the type of spatial arrangement used has minimum impact on detectability of rare species (Berec et al., 2015). To maximise detectability, we placed microphones at a height of approximately 1.5 m above the ground and orientated them horizontally and away from vegetation (Weller & Zabel, 2002). We checked microphone sensitivity prior to the start of the survey period, as well as approximately halfway through, using an ultrasonic testing device (Ultrasonic Calibrator; Wildlife Acoustics) with a 40 kHz pulse. Data obtained from nearby weather stations indicated that conditions complied with best practice guidelines for bat surveys in the UK (sunset temperature ≥ 10 °C, no rain or strong wind; Collins, 2016) on all study nights (see Article S2 for recorded conditions).

Bat calls were distinguished from background noise using SonoBat software (version 4.5.0, SonoBat, Arcata, CA, USA). All remaining files were then subjected to manual checking, with reference to parameters provided in Russ (2012). We considered a file to be equivalent to a single bat pass (traditionally defined as a call or series of calls, separated by no more than a one second time gap; Fenton, 1970), as preliminary analysis indicated that multiple passes of barbastelles occurred in only 2.7% of files. As barbastelles recorded within the first hour after sunset fall within the species-specific emergence time range (Russ, 2012), passes close to dusk may also be indicative of a nearby roost. Therefore, we conducted analysis on activity recorded within the first hour after sunset. Whilst high activity close to dawn may be indicative of a nearby roost, we did not analyse this time period as previous work found return times to be highly variable among barbastelle individuals ($\bar{X}=$ 194 ± 59.1 min before sunrise; Zeale, Davidson-Watts & Jones, 2012), whereas emergence times were much more consistent.

Woodland coverage maps

A given unit increase in detector number has a smaller impact on detector density as woodland size increases. There is also likely to be more heterogeneity of habitat in larger parcels compared with smaller ones. As such, there is a need to standardise surveys to account for the spatial variation in activity, as well as the size of a site. We therefore estimated detector density by constructing smoothed woodland coverage maps by fitting concave hulls (alpha hulls) to detector locations in QGIS (version 3.6.3-Noosa; QGIS Development Team, 2020). The alpha hull is an algorithmic method for assigning a boundary around a discrete set of points. They are constructed by creating a Delaunay triangulation of all points within the sample and subsequently retaining only those vertices which are shorter in length than the chosen value of the parameter alpha (Burgman & Fox, 2003). Varying the value of alpha between zero (i.e., a set of discrete points) and one (i.e., a minimum convex polygon) generates different hull configurations, which will include increasingly isolated points as the value increases. In this study, we set the value of alpha to 0.7, representing the best balance between the inclusion of all points and the exclusion of areas without any detectors (for additional methodological details, see Fig. S1 and Article S3). We used alpha hulls rather than minimum convex polygons to minimise spatial bias and overestimation, which are particularly problematic for range and distribution estimates (Burgman & Fox, 2003).

Phase 2: Application of methodology

We undertook acoustic bat surveys in four regions across England between May and September in 2021 (Fig. 1). Each region consisted of a 35 × 20 km area, in the counties of West Sussex, Wiltshire, Herefordshire, and the South Midlands (Northamptonshire, Buckinghamshire, Bedfordshire). The regions represented contrasting agricultural landscapes, and were all within the core geographical range of the species.

Within these regions, we selected study sites from broadleaved and mixed woodland in the National Forest Inventory. A stratified random sampling approach was used, using woodland patch size as the strata. Most woodland patches in England are small, and therefore simple random selection of study sites would have yielded a sample with few large woodlands, preventing any possible future investigation of the association between patch size and probability of occupancy. In addition, woodland size may be a constraint on barbastelle colony formation. This has been demonstrated for the Bechstein’s bat (Myotis bechsteinii), a rare woodland bat with similar woodland structural requirements, which has been shown to need around 70 ha of good quality habitat to support colonies of 20–30 females in the Southern Upper Rhine region (Steck & Brinkmann, 2013). We therefore defined the minimum size category based on the smallest woodlands with barbastelle records, according to the most comprehensive available national assessment (Mathews et al., 2018). The remaining three categories were based on equal count quantiles of all deciduous woodland that had barbastelle records: (i) very small ≤ 1.04 ha, (ii) small > 1.04–18.48 ha, (iii) medium > 18.48–68.67 ha, and (iv) large > 68.67–1779.29 ha. We randomly selected five woodlands from each size category per region, giving a total of 80 sites. If we could not obtain landowner permission for a given woodland, the nearest accessible woodland that fell within the same size category was selected.

Volunteers were recruited from Spring 2021 to participate in the Barbastelle Volunteer Bat Survey (https://www.vwt.org.uk/research-all/monitoring-for-the-near-threatened-barbastelle/) and provided with pre-set acoustic detectors (Wildlife Acoustic SM2 and SM2+ detectors recording in full-spectrum at either 192 or 384 kHz, respectively). The number of units provided was dependent on the size of the woodland to ensure approximate equal detector density across sites. Both the density of detectors used and the duration of the survey were determined based on the findings of the methodological development phase. Volunteers were instructed to place detectors along trails or rides (unless none were available) in the pre-selected woodlands. We provided a GPX file that included the recommended location of each detector, allowing volunteers to download and view the survey locations on a mobile phone, and asked them to maintain an approximately even distribution of detectors through the woodland if they had to deviate from the recommended locations. Volunteers recorded the grid reference and unique ID number of each detector on a recording form. For each survey, we used the same microphone placement, detector settings, acoustic analysis, and minimal weather conditions as adopted during the development phase. Once the survey was completed, volunteers returned all equipment back to one of the survey bases for analysis. We provided volunteers with feedback in the form of a detailed survey report that summarised the bat species and barbastelle activity recorded at each detector.

Phase 3: Validation by radio-tracking

We undertook trapping surveys in 2022 within a subsample of the woodlands acoustically surveyed by volunteers in 2021. Initially, we separated all woodlands into two categories (high and low probability of colony presence) based upon the optimal cut-off threshold identified in the development phase (see statistical analyses, below). Trapping was conducted in one randomly-selected low probability woodland in each of the four study regions. A further 13 high-probability woodlands were surveyed, with woodlands distributed approximately evenly between the study regions. Due to time availability within the season and logistical constraints, we prioritised surveying high-probability woodlands as our primary aim was to assess whether maternity colonies could be located using our methodology.

Nets were set at the detector location that recorded the highest number of barbastelle passes (defined as the highest number of barbastelle passes recorded within an hour of sunset, across all nights a detector was deployed). If there were no barbastelles passes at any detector within this time period (occurred only in the low-probability woodlands), then the location with the most barbastelle passes across the entire night was used. When we recorded no barbastelles within a woodland across the entire night, then the location was randomly selected from amongst the detector locations. If a location was deemed unsuitable (i.e., due to accessibility and practicality) then we selected the next suitable location based on the criteria described above.

A combination of monofilament mist nets (total length = 45 m; Ecotone, Poland) and two harp traps (Faunatech Austbat, Victoria, Australia) was used at each site. A Sussex Autobat lure (Autobat Mk 2; Autobat, Sutton Coldfield, UK) broadcasting barbastelle echolocation and social calls, either in isolation or in sequences including other species, was used in conjunction with harp traps. Nets were opened for two hours from sunset and each location was surveyed for two nights. Adult female or juvenile barbastelles were fitted with lightweight radio transmitter PicoPip (Ag377, 0.29 g; LoTek, Ontario, Canada) or Holohil (LB-2X, 0.27 g; Holohil, Ontario, Canada) tags weighing < 5% of the bats weight. We clipped the fur of each bat and attached a tag between the scapulae using a flexible latex-based glue (either Ostomy Adhesive Salts Healthcare Ltd., Birmingham, UK, or Torbot Bonding Cement, Torbot Ltd, Cranston, RI, USA). We tagged a maximum of 2–3 individuals at each woodland site, and priority was given to parous females, followed by juvenile animals (identified by the lack of fused epiphyses), given the objective of identifying maternity roosts.

Tagged bats were tracked to roost trees the following day using Australis or Regal (Titley Electronics, New South Wales, Australia), Sika (Biotrack Ltd, Wareham, United Kingdom), or Biotracker (Lotek, Newmarket, Canada) VHF receivers with three-element Yagi antennas. We used an omni-directional antenna mounted on a vehicle to assist with tracking outside study woodlands. Possible occupied tree cavities were identified via visual inspection. Emergence surveys were then conducted, starting at sunset and lasting for 1.5 h or until bats had finished emerging. Video recordings to permit assessment of the number of bats emerging were made using Canon XA20 and XA40 infrared cameras (Tokyo, Japan), mounted on tripods and positioned on either side of the tree, in conjunction with two infrared illuminators (LIR-IC88, 850 nm; IRLAB, Shenzhen, China; range = 180 m; beam angle of 40°).

Statistical analysis

We conducted statistical analyses in RStudio using R version 3.6.1 (R Core Team, 2019) and produced graphical outputs using the ggplot2 package (Wickham, 2016). The residuals of each model were checked for normality and heteroscedasticity using the ‘testResiduals’ function in the DHARMa package (Hartig, 2022).

For phase 1, we analysed the activity within the first hour after sunset in order to identify a cut-off threshold useful for distinguishing woodlands with and without barbastelle colonies. Where the full hour after sunset was not recorded completely, owing to occasional detector failure (approximately 5% of deployments—24 detector-nights in total), it was excluded. We took the maximum activity recorded across all the detectors within an hour of sunset to represent the best indication as to whether a given site was occupied by a colony. For each night, we selected the detector recording the highest level of activity for each site where the presence or absence of colonies was known (i.e., excluding sites categorised as unknown).

We analysed the relationship between the pass rate and the presence/absence of a colony using a generalised linear model (GLM), specifying a binomial error structure with a logit link function. Colony presence was the response variable, bat passes within the first hour after sunset was the predictor, and each woodland was treated as a replicate. To determine the predictive power of the model and to calculate the optimal cut-off threshold for bat passes, we derived a receiver operating characteristic (ROC) curve using the package pROC (Robin et al., 2011). ROC statistics are informative in evaluating binary class decisions and cut-off parameterisation based on a models’ true-positive (sensitivity) and false-positive rate (1 - specificity) at various threshold values. We used the ‘coords’ function in pROC, specifying “best” as an argument, to select the threshold corresponding to the best sum of sensitivity and specificity respectively.

To determine the survey effort (i.e., detector density) required to identify the presence of colonies during phase 1, we constructed power curves using a random subsampling with replacement approach. For each site with a known or predicted colony, we calculated the density of detectors per smoothed alpha hull, then incrementally eliminated detectors randomly. Each time a detector was removed, the density was recalculated along with the maximum number of barbastelle passes that were recorded from the new subset. We compared this value with the optimal cut-off threshold derived from the ROC curve generated from the entire dataset, and categorised woodland as likely to contain a colony (value of 1) or not (value of 0). We estimated survey sensitivity (i.e., probability of detecting a colony by at least one sample) as the proportion of 1,000 stochastic simulations in which at least one static detector had pass rates higher than the ROC-derived threshold.

We modelled the relationship between detector density and survey sensitivity using a generalised additive mixed model (GAMM), specifying a binomial error structure (mgcv package; Wood, 2011). Detector density was specified as a predictor, fitted with penalised thin plate regression splines (knots = 3), and survey night was specified as a random effect. We used model predictions to determine the detector density required to maintain a specified minimum survey sensitivity. Following the precautionary principle, to minimise the possibility of concluding that no colony was present when in fact there was one (false-negative, Type II error), we adopted a stringent survey sensitivity of 0.9 (Jones, 2013; Niver et al., 2014; Meyer, 2015).

To determine the appropriate survey duration to be used in phase 2, we analysed the effect of duration on the prediction of colony presence. We conducted this analysis for two sites where we had complete acoustic data sets for five consecutive nights. For the full range of detector densities, we calculated the survey sensitivity for every possible combination of nights that could have been surveyed (1–5 nights). We modelled this relationship using a generalised additive model (GAM), specifying a binomial error structure (mgcv package; Wood, 2011). Survey sensitivity was used as our response variable, whilst detector density and number of survey nights (both fitted with penalised thin plate regression splines - knots = 3) were specified as predictors.

For phase 3, the relationship between pass rate and capture success was analysed using a generalised linear model (GLM), specifying a binomial error structure with a logit link function. Capture success of pregnant/lactating or juvenile barbastelle bats (i.e., bats potentially suitable for radio-tracking to a maternity roost) was defined as the response variable, and the maximum bat passes recorded in that woodland during the citizen science surveys was defined as the predictor. ROC analysis was subsequently conducted, in the same manner as previously described, to determine the optimal cut-off threshold for bat activity (pass rate) likely to yield successful capture of barbastelle bats.

Ethical approval

The work was approved by the Animal Welfare and Ethical Review Committee of the School of Life Sciences, University of Sussex (Ethical Application Ref: ARG-25). It was also part of work conducted under UK Home Office Licence number PIL P57B69020 and Natural England Licence no. 2022-61108-SCI-SCI and preceding licences.

Phase 1: Acoustic methodology development

We monitored bat activity for 470 detector-nights across 13 woodlands during the summer of 2019 (Table 1). Throughout the study, 1,549 barbastelle passes were recorded within the first hour of sunset, occurring at a mean time of 35 min (SD = 11.6) following sunset.

Of the 13 woodlands surveyed, historical evidence identified six sites as supporting colonies, whilst four sites were classified as having no colonies present. For the remaining three sites, we were unable to determine whether they were occupied or unoccupied by colonies based on our analysis of temporal activity patterns, and therefore these sites were classified as ‘unknown’ in our analysis. Among the sites without colonies, three exhibited minimal barbastelle activity throughout the entire survey period (range: 8–15 passes), with no clear peaks in activity. The remaining site recorded a substantial number of barbastelle passes (n = 180), but very few within the first hour across the survey period (n = 9); corresponding with findings from a recent Environmental Impact Assessment which concluded that the site was used by barbastelles for foraging but not breeding (Highways England, 2018).

The number of bat passes within an hour of sunset strongly predicted the probability of colony presence or absence (GLM: OR = 3.28, 95% CI [1.24–8.65], z_1,33 = 2.40, p = 0.018). Follow-up ROC analysis (AUC = 0.966, 95% CI [0.898–1]) identified that a cut-off point of 3.45 barbastelle passes within an hour of sunset optimises the balance between true-positive (sensitivity) and false-positive rates (1 - specificity) in predicting colony presence (Fig. 2). Using this value, 95.7% of survey nights in woodlands with known colonies, and 100% of survey nights where a colony was absent from the woodland, were correctly predicted. To accommodate practical considerations, this threshold is rounded up to four passes, as having a fraction of a pass is not possible. Subsequently, two woodlands that were previously unknown to contain colonies were identified as likely to support colonies (passes < 1 hour of sunset; site 11 = 6, site 13 = 18).

There was a significant non-linear relationship between increasing detector density and the probability of woodlands being correctly classed as positive for barbastelle roost occupancy (GAMM, Figs. S2 and S3). The effect of detector density on woodland classification was highly significant for each woodland (p < 0.001), with the survey effort required to reliably detect a colony varying depending on the woodland surveyed. Overall, there was a 90% chance of a woodland being correctly identified as having a colony with a density of 0.16 detectors ha⁻¹ (Fig. 3). The adjusted-R² value for the model was 0.936. When only a single detector within a woodland was selected via our random subsampling approach, we found that there was just a 32% chance of detecting a colony based on the average survey effort. Survey duration had minimal effect on the ability to detect a colony, based on a survey sensitivity of 0.9 (Fig. 4). Each site showed similar detector densities needed to detect colonies, regardless of survey duration (GAM prediction ranges: site 1 = 0.106 − 0.107 detectors ha⁻¹, site 5 = 0.175 − 0.177 detectors ha⁻¹).

Phase 2: Application of methodology

Based on the results of phase 1, volunteers undertook surveys using a minimum detector density of 0.16 detectors ha⁻¹ for a duration of three consecutive nights. Whilst the results of phase 1 indicate that a shorter survey duration may be sufficient to detect a colony, we chose three nights to account for the statistical uncertainty of our results due to a limited sample size. Surveys were completed at 509 different locations across 77 woodlands (three woodlands could not be surveyed due to time constraints). This comprised 1826 complete nights of recording, which resulted in the collection of 1,666 barbastelle recordings within the first hour after sunset. Barbastelles were recorded within one hour of sunset in 40 woodlands (52%), of which 26 (34%) were predicted to support colonies based on the optimal cut-off threshold identified in phase 1 (Fig. 5).

Phase 3: Validation by radio-tracking

Of the 26 woodlands predicted to support colonies, trapping surveys were conducted at 13 and barbastelles captured at 10. Trapping was also conducted at four sites that were predicted to have no colonies, and barbastelles were captured at one of these (a single adult male). Radio-tracking of 11 adult females and two juveniles in eight woodlands (those with bats suitable for tagging) led to the identification of five previously unidentified maternity colonies. Multiple roost trees were identified in two woodlands, though in both cases these were assumed to be part of the same colony due to observed roost-switching and proximity. All colonies were located within the woodlands where the bats were captured, and roost counts ranged from 14 to 52 bats (Table 2; $\bar{X}=$ 30.60, SD = 15.55). Out of the seven tree roosts identified, four were found under exfoliating bark of standing dead trees. A further two roosts were identified within live trees - one within a shearing crack and the other in a crevice formed by a tree limb breaking off. The feature used for the remaining roost could not be identified as the three radio-tagged bats switched to a second roost between nights, and a further emergence survey could not be conducted due to time constraints. Of the woodlands predicted to have barbastelle colonies, barbastelles suitable for tagging (adult females or juveniles) were caught in 69% of sites, and the colony’s roosting location was located in 38%.

The number of bat passes recorded within an hour of sunset was a strong predictor for the successful capture of bats associated with maternity colonies (adult females or juveniles) (GLM: OR = 1.24, 95% CI [1.11–1.44], z_1,71 = 3.35, p < 0.001). The ROC analysis, based on the radio-tracking data, indicated that there would be a 33% chance of capturing barbastelles potentially suitable for radio-tracking back to colonies if three bat passes were recorded within the first hour after sunset at a detector (AUC = 0.856, 95% CI [0.754–0.9581]). When 16 or more bat passes were recorded then the probability of capture success increased to 90%.