Tradeoffs in sound quality and cost for passive acoustic devices

ABSTRACT

Ecologists often collect data with automatic sensors such as passive acoustic devices. Traditional commercial passive acoustic devices that record ultrasonic sounds are usually > 700USD; less costly options now exist but the sound quality is unknown. We compared a low-cost passive acoustic device to a traditional device for quality of bat call data recorded. We simultaneously deployed two device types at seven sites to record free-flying bats over three nights. We paired call files from both devices when they occurred ≤ 3 s apart and contained ≥ 3 pulses; for each file pair and pulse, we measured indicators of recording quality, specifically duration, maximum frequency, and bandwidth. We compared the proportion of files classified as low frequency, mid frequency, or Myotis bats. The low-cost device recorded lower quality pulses (i.e., shorter duration, lower maximum frequency, and shorter bandwidth). The more expensive device recorded a higher proportion of Myotis calls and a lower proportion of low-frequency calls, but there was no difference across devices in the proportion of calls classified as mid-frequency bats. Less expensive alternatives for passive acoustic devices may be inappropriate for studies requiring high-quality data but may be critical when research objectives call for multiple devices and are constrained by budget.

KEYWORDS:

• Bandwidth
• bats
• duration
• frequency
• passive acoustic monitoring
• sensor cost

Introduction

Large-scale studies help us understand how regional or even global threats, such as human impacts, affect ecosystems of interest (Ahumada et al. 2011; Wrege et al. 2017). By monitoring multiple sites simultaneously, we gain a more accurate picture of the community, population size, or animal movements compared to monitoring multiple sites asynchronously. However, to detect rare species and increase the probability of detecting all species in an area, we need to sample more sites on the landscape (MacKenzie and Royle 2005; Skalak et al. 2012; Banner et al. 2019). With limited budgets, it may be impossible to simultaneously deploy sufficient numbers of devices to achieve desired fine-scale spatial and temporal precision. Less expensive devices would help overcome this issue.

Ground-level sensing technologies are beneficial for large-scale studies, even when remotely sensed data are available (Peres et al. 2006). There are many types of automated sensors – e.g. for recording sounds, photos, weather conditions, chemicals, locations, and other parameters. For example, Tempa et al. (2019) deployed camera traps in 1,129 grid cells to survey tigers across a 38,000-ha landscape, enabling them to track tiger movements and to estimate population density. Rodhouse et al. (2015) deployed passive acoustic devices at 241 sample units across the states of Washington and Oregon in the United States to establish baselines of occurrence levels for bat populations. Passive acoustic monitoring (PAM) uses automatic sensors to record sound in the environment in the absence of a human observer. PAM is a technique used to record bats, nonvolant terrestrial mammals, anurans, birds, marine animals, insects, and soundscapes (e.g., Palmer et al. 2019; Sugai et al. 2019). Myriad devices exist for PAM, with variations on detection ranges, sensitivity to different sounds, weatherproofing, cost, and other characters (Fraser et al. 2020). Multiple passive acoustic devices set up near each other (arrays) can be used to determine locations of animals, such as marine mammals or bats, in a two-dimensional (Weller and Baldwin 2012) or three-dimensional space (Koblitz 2018; Gillespie et al. 2020).

Bats are difficult to detect visually, but relatively easy to detect by sound, so bats are often monitored via PAM. Many bats emit high-frequency sounds (typically >20 kHz) and interpret the incoming echoes to identify targets and for navigation (echolocation). A bat echolocation call is made up of pulses of sound, and key characters for the identification of pulses include duration (milliseconds, ms), frequency at different points in the call (kilohertz, kHz), and frequency bandwidth (kHz). Echolocation frequencies vary widely among bat species and can even vary within one species according to geographic location, proximity to water, vegetative clutter, or the presence of conspecifics (Fenton and Bell 1981; Thomas et al. 1987; Kalko and Schnitzler 1993; Obrist 1995). For example, the short-eared trident bat (Cleotis percivalis) can echolocate up to 212 kHz and the spotted bat (Euderma maculatum) echolocates as low as 8 kHz (Fenton and Bell 1981). Generally, smaller bats echolocate at higher frequencies (Bogdanowicz et al. 1999). Recording those very high frequency sounds is important to the study of bats. In the eastern United States, federally threatened and endangered bat species generally echolocate with maximum frequencies >50 kHz (Murray et al. 2001). To adequately survey for these species using PAM requires devices that are capable of detecting and recording adequate representations of sounds emitted within a reasonable sampling area.

Bats echolocate at high amplitudes, but their high-frequency sounds attenuate rapidly compared to low-frequency sounds (Griffin 1971). Furthermore, high-frequency sounds attenuate even more rapidly in certain conditions – e.g. hot, humid weather (Griffin 1971; Lawrence and Simmons 1982; Goerlitz 2018) and where there are high levels of vegetative clutter near the microphone (Patriquin et al. 2003). Microphones must be designed specifically to detect ultrasonic frequencies (Parsons and Szewczak 2009), but microphones from different manufacturers have varying sensitivities and do not record bat echolocation calls equally (Adams et al. 2012; Smith et al. 2020). Microphones can be omnidirectional or directional, which impacts what is recorded. Omnidrectional microphones have a more rounded sphere of detection around the microphone and a shorter detection range than directional microphones (Limpens and McCracken 2004). Along a suspected flight path for bats, such as a forest edge, directional microphones can be focused to record towards the flight path, while an omnidirectional microphone will record sounds from along the flight path and in the vegetation, which may yield more non-target sounds (Limpens and McCracken 2004). Kaiser and O’Keefe (2015) showed that weatherproofed omnidirectional microphones recorded more high-frequency echolocation calls, whereas weatherproofed directional microphones recorded more low-frequency echolocation calls.

Although large-scale studies are beneficial in ecology, the high cost to collect data over a large area could impair sampling efforts (Jones 2011). Some acoustic devices for bats cost > 1,000USD (e.g. Anabat Swift, Wildlife Acoustics SM4BatFS, Pettersson D500×, Avisoft Ultrasoundgate) per unit, making it difficult for researchers to obtain enough devices to conduct large-scale studies. High cost may be even more of a burden for researchers in countries with fewer economic resources (e.g. Mammides et al. 2016), even though some of these countries may harbour high levels of biodiversity and are targets for conservation research (Wilson et al. 2016). Securing expensive devices to protect from theft, weather, and wildlife further increases costs. The loss of an expensive device can set a research study back, especially if the device is difficult to obtain. Less expensive devices would make it more cost effective to implement large-scale conservation ecology studies. However, switching to a less expensive device may require a tradeoff between cost and quality.

Our goal was to compare the quality of recorded echolocation calls of free-flying bats for a widely used, commercial passive acoustic device that is > 1,000USD and a novel, open-source device that is < 100USD. Our specific objectives were to compare the two devices with respect to 1) the proportion of echolocation calls categorised into one of three frequency ranges – i.e. phonic groups and 2) call parameter measures indicative of quality. We predicted that, for each of three phonic groups, the less expensive device would record fewer files than the more expensive device. We also predicted that the less expensive device would record echolocation calls of lower quality (shorter duration, lower maximum frequency, and shorter bandwidth) than the more expensive device.

Materials and methods

We deployed acoustic detectors from 19–21 July 2019. We recorded echolocation calls at seven randomly selected sites within the Atlanta-Long Branch Conservation Area in Macon County in northeastern Missouri, USA. To reduce bias from spatial autocorrelation or high bat activity on waterbodies, sites were ≥500 m from another site, >50 m from non-ephemeral waterbodies, and >100 m from a flowline crossing the corridor. Flowlines and waterbody areas were delineated in pre-existing shapefiles (National Hydrography Dataset, USGS) or digitised in Google Earth (version 7.3.3). Weather conditions at the nearby (approx. 25 km away) Kirksville Regional Airport on 19, 20, and 21 July were the following: mean temperature = 30.9, 29.1, and 24.3°C; maximum relative humidity = 75, 94, and 100%; precipitation = 0, 12.2, and 1.5 mm, and mean wind speed = 5.1, 5.1, and 3.5 ms⁻¹; respectively (National Weather Service data obtained from the Midwestern Regional Climate Center, cli-MATE). Although there was rain at the weather station on nights two and three, there was no evidence of rain accumulation at the recording sites. Bat species in the area are Indiana bat (Myotis sodalis), little brown bat (M. lucifugus), northern long-eared bat (M. septentrionalis), big brown bat (Eptesicus fuscus), eastern red bat (Lasiurus borealis), hoary bat (L. cinereus), silver-haired bat (Lasionycteris noctivagans), evening bat (Nycticeius humeralis), and tri-coloured bat (Perimyotis subflavus). Echolocation frequencies for these bats range from a minimum of 15 kHz (hoary bat; O’Farrell et al. 2000) to a maximum of >100 kHz (northern long-eared bat; Broders et al. 2004). For analysis, as is standard for automated identification software, we used the average characteristic frequency (measured at the flattest portion of a call) to assign call files to low-frequency (<30 kHz), mid-frequency (30–45 kHz), and Myotis (>45 kHz) phonic groups (Britzke and Murray 2000; Clement et al. 2014). This coarse classification scheme mitigates for the difficulty of reliably differentiating species with overlapping echolocation call characteristics (Rydell et al. 2017) and considers foraging habits. Bats in the low phonic group generally forage in open space (big brown bat, hoary bat, and silver-haired bat); bats in the mid phonic group usually forage along edges (eastern red bat, evening bat, and tri-coloured bat), and bats in the Myotis phonic group tend to forage in forest interior (Indiana bat, little brown bat, and northern long-eared bat; e.g. Beilke et al. 2021).

We tested the AudioMoth (version 1.1.0) and Anabat Swift. The AudioMoth is a low-cost (70–86USD for device + internal microphone) passive acoustic device (Open Acoustic Devices). The Swift is a higher cost unit (1274USD for device + external microphone; Titley Scientific). Both devices are full-spectrum recorders that can be programmed to record during a certain time frame and at certain frequencies. The AudioMoth has dimensions of 58 × 48 × 15 mm, which makes it much smaller than the Swift (182 × 119 × 43 mm). For the Swift, we used a directional microphone that was 60 mm long and 46 mm in diameter, oriented parallel to the ground and directed towards an open field or trail along a forest edge. The AudioMoth microphone is flush with the face of the device and is omnidirectional. The AudioMoth recorded all sounds within the user defined recording period. The Swift was constantly listening to sounds but only recorded a sound to a file when a specified minimum frequency and minimum sound duration triggered the device.

We weatherproofed each AudioMoth by stretching a plastic bag (0.05 mm thick) over the unit until taut and then sealing with minimal air inside (Hill et al. 2018). We constructed weatherproofing for the stainless-steel Swift microphones (Titley Electronics 2019) with seal foam weatherstrip securing the microphone inside a 4.4-cm polyvinyl chloride (PVC) tube with an end cap suspending the microphone from its base. The main unit was waterproofed by the manufacturer.

We mounted each weatherproofed Swift microphone onto a 3-m high PVC pole set along a forest edge ≥12 m from the next closest forest edge and ~2 m away from trees behind the microphone. Each AudioMoth was attached to the pole with zip ties about 15 cm below each Swift microphone. We used a 10-m cable to connect the Swift microphones to the detectors, which were secured to a tree. The microphone can be directly attached to the Swift, but to adequately weatherproof the microphone and avoid placing the larger unit at the top of the pole, we chose to use a cable connection; however, the cable may attenuate sound recordings.

We deployed the two devices simultaneously for three consecutive nights, programming them to turn on 30 min prior to sunset and to turn off 30 min after sunrise each night (20:04 to 06:27 on July 19, 20:04 to 06:28 on July 20, and 20:03 to 06:29 on July 21). We used the AudioMoth Configuration App (Open Acoustic Devices 2019) for programming the AudioMoths with the following settings: recording period = 5 s, rest period = 4 s, sampling rate = 384 kHz, and mid gain. We programmed the Swifts with the following settings: constant recording = off; minimum frequency = 16 kHz; maximum frequency = 250 kHz; minimum event = 3 ms; trigger window = 2 s; sampling rate = 500 kHz; max file length = 10 s, Analog HP filter = on; sdfilename prefix = off. We selected these settings to align with prior work (Roemer et al. 2017; Millon et al. 2018; Kemp et al. 2019; López-Baucells et al. 2019). We powered AudioMoths with three AA rechargeable batteries and recorded calls to a 16-GB microSD card. The Swifts were powered with four AA rechargeable batteries and recorded to two 32-GB SD cards.

The AudioMoth detectors named files using a hexadecimal format, so we used program R (base R; functions: ‘as.hexmode’, ‘as.integer’, ‘as.POSIXct’, ‘dir’, ‘gsub’, ‘length’, ‘paste0’, ‘substr’, and ‘file.rename’; version 3.6.2, R Core Team 2019) to rename the files with a site identifier and a date and time stamp. We ran all Swift and AudioMoth recorded calls through Bat Call Identification (BCID) autoclassification software (version 2.8b, Kansas City, MO, USA; Allen 2019), so we converted all full-spectrum data to zero cross files in Anabat Insight (version 1.9.0–5-g38ddee8, Titley Scientific, Brendale, QLD, Australia), because BCID does not accept full-spectrum files. Amplitude information from full-spectrum files may be beneficial to bat identification, but it can also be sensitive to many variables not under our control (Agranat 2012). We used the default filter in BCID to identify recordings as a bat call. The program automatically assigned each call to one of five categories, three of which were the assigned phonic groups: low frequency, mid frequency, Myotis (bats classified in the genus Myotis), unknown, or noise. Unknown (<1% of files) and noise files were discarded. We visually examined the calls identified as low, mid, and Myotis bat calls to confirm that they were indeed bat calls.

To compare the quality of recorded files across the two detector types, for each site and night, we created a subset of AudioMoth files and Swift files that were recorded within a ≤ 3s window and for which the AudioMoth file in the pair was identified as a bat call by BCID (n = 890). We used Anabat Insight to visually examine the recording pairs to look for consistency in characteristic frequency, slope (the change in pitch over time; octaves per second), and time between calls within the recording pairs. We omitted calls if they appeared to contain pulses from multiple species or phonic groups. We only kept pairs in the analysis where both units recorded ≥ 3 pulses in a file. We chose 3 pulses as a cut-off to ensure that a bat was recorded and to minimise identification errors (Ford et al. 2005). For the file pairs retained, we used the generate report feature in the analysis tab in Anabat Insight to automatically generate a metrics report which gave us the number of pulses, mean duration of the call (ms), mean maximum frequency (kHz), and mean minimum frequency (kHz) for each call file. We calculated the frequency bandwidth (kHz) by subtracting the mean minimum frequency from the mean maximum frequency. We discarded files with unreasonable values where a program error had occurred (e.g. negative numbers or numbers > 10000 for frequencies; n = 32). We focused on duration, bandwidth, and maximum frequency to compare call quality between the two types of detectors.

For this subset of file pairs (n = 858), we compared linear mixed-effect models in program R (version 4.0.3; package lme4, Bates et al. 2015) with fixed effects of detector and date, and random effects of file pair and site on each of these response variables: duration, maximum frequency, and bandwidth. We log-transformed the response variables to normalise the data. We included date as a fixed effect to account for weather effects. Models tested included 1) detector, 2) date, 3) date + detector, 4) date × detector to assess differential effects of weather by detector type, and 5) a null model. All models included file pair and site as random effects.

To compare data acquisition by phonic group, for each detector type and site-night combination, we divided the nightly sum of call files for each phonic group by the nightly sum of all files classified as a bat call. These proportions ranged from 0 to 1. We transformed these proportional values to be > 0.0 and < 1.0 (Smithson and Verkuilen 2006) and used beta regression models in program R (version 4.0.3; package betareg, Cribari-Neto and Zeileis 2010). For each phonic group (low frequency, mid frequency, and Myotis), we compared four models to test the fixed effects of detector and site on the proportion of all calls identified to that phonic group. The models tested included 1) detector, 2) site, 3) detector + site, and 4) a null model.

For all model comparisons, we ranked model by AICc scores, considering models plausible if ΔAICc was ≤ 2 units from the top model (Burnham and Anderson 2002). We considered the fixed effects significant if 85% confidence intervals did not cross zero (Arnold 2010).

Results

At seven sites, the AudioMoth devices were on for a total of 171.80 h and recorded 76,196 data files during the scheduled recording times. Of all the files that contained recorded sound, 2,479 (3.3%) were identified as containing a bat call. Simultaneously, the Swifts were actively listening for bat sounds and recorded 6,727 files over 187.57 h, and 5,063 (75.3%) files were classified as bat calls. We retained a subset of 890 file pairs recorded within 3 s of each other by either device.

The date × detector model was the best for predicting duration, bandwidth, and maximum frequency (model weights = 0.83, 0.93, and 0.97, respectively; Table 1). When compared with AudioMoth files, Swifts recorded calls of significantly higher quality (i.e. longer duration, greater bandwidth, and higher maximum frequency). Swifts recorded echolocation calls 24% longer in duration, 63% greater in bandwidth, and 16% higher in maximum frequency (Table 2 and Figure 1). However, the effect of detector was moderated by date (Table 2 and Figure 1), suggesting that weather affected quality differences. The recording conditions were different on the second and third nights when humidity was higher; further, there was a nearby storm on the second night.

Figure 1. Duration (ms), bandwidth (kHz), and maximum frequency (kHz) of bat calls recorded by AudioMoths and Anabat Swifts by date in northeast Missouri in July 2019.

Table 1. Models used for each call characteristic (duration, bandwidth, and maximum frequency) to compare between AudioMoth and Anabat Swift acoustic detectors. For each model, we present the degrees of freedom (df), ΔAICc, model weight, and conditional R² values.

Model	df	ΔAICc	weight	R^2
Duration (log Duration)
Date × Detector	9	0.00	0.83	0.67
Date + Detector	7	3.51	0.14
Detector	5	7.21	0.02
Date	6	536.65	0.00
NULL	4	540.35	0.00
Bandwidth (log Bandwidth; Fmax-Fmin)
Date × Detector	9	0.00	0.93	0.74
Date + Detector	7	5.88	0.05
Detector	5	7.41	0.02
Date	6	959.10	0.00
NULL	4	960.63	0.00
Maximum Frequency (log Max Frequency)
Date × Detector	9	0.00	0.97	0.76
Detector	5	7.56	0.02
Date + Detector	7	10.36	0.01
NULL	4	837.36	0.00
Date	6	840.16	0.00

Table 2. Estimates and 85% confidence intervals (CI) for independent linear mixed effect models comparing call characteristics of bat calls collected by Anabat Swifts and AudioMoths. AudioMoth is used as the reference variable for Detector and 19 July is used as the reference variable for Date. Site was a random effect in each model.

Model	Fixed Effect	Estimate	85% CI
Duration ~ Date × Detector	Intercept	1.278	1.206	1.351
	Detector (Swift)	0.234	0.216	0.251
	Date (July 20)	0.052	0.021	0.081
	Date (July 21)	−0.016	−0.044	0.013
	Detector (Swift) × Date (July 20)	−0.053	−0.081	−0.025
	Detector (Swift) × Date (July 21)	−0.021	−0.047	0.005
Bandwidth ~ Date × Detector	Intercept	2.060	1.931	2.188
	Detector (Swift)	0.489	0.463	0.516
	Date (July 20)	−0.008	−0.054	0.037
	Date (July 21)	0.008	−0.035	0.051
	Detector (Swift) × Date (July 20)	−0.050	−0.092	−0.008
	Detector (Swift) × Date (July 21)	0.046	0.006	0.085
Max frequency ~ Date × Detector	Intercept	3.750	3.711	3.789
	Detector (Swift)	0.161	0.152	0.170
	Date (July 20)	0.010	−0.007	0.028
	Date (July 21)	0.009	−0.008	0.026
	Detector (Swift) × Date (July 20)	−0.037	−0.051	−0.023
	Detector (Swift) × Date (July 21)	−0.010	−0.023	0.004

When we considered the full suite of calls identified to a phonic group for both types of detectors, we observed variation by site and detector type for two of three phonic groups. The detector + site model was the best for predicting the proportion of low and Myotis calls, but site only was best for predicting the mid-frequency group (Table 3). Although the simpler model (site only; ΔAICc = 1.77) was also plausible for the low frequency phonic group, we interpreted the model that included detector, because we were more interested in the effects of detector than site. The proportion of bat calls that were identified to the mid-frequency phonic group was similar between the AudioMoth and Swift. However, the Swift recorded a significantly higher proportion of Myotis bat calls (10% of all bat call files vs. 5% for the AudioMoth) and a lower proportion of low-frequency calls (18% vs. 25% for the AudioMoth; Table 4 and Figure 2).

Figure 2. Proportion of all AudioMoth and Anabat Swift echolocation recordings (n = 2,049 for AudioMoths and 5,074 for Anabat Swifts) classified as low frequency, mid frequency, and Myotis bats from calls recorded at seven sites in northeast Missouri, July 2019. The value for each site is the average over three nights of recording.

Table 3. Models used for comparing proportion of calls identified to each phonic group by Anabat Swift and AudioMoth acoustic detectors. For each model, we present the degrees of freedom (df), ΔAIC, model weight, and pseudo R² values.

Model	df	ΔAICc	weight	R^2
Low Frequency
Detector + Site	9	0.00	0.71	0.66
Site	8	1.77	0.29
NULL	2	30.18	0.00
Detector	3	30.92	0.00
Date	4	31.85	0.00
Detector + Date	5	32.53	0.00
Mid Frequency
Site	8	0.00	0.84	0.66
Detector + Site	9	3.25	0.16
NULL	2	34.81	0.00
Detector	3	37.14	0.00
Date	4	37.38	0.00
Detector + Date	5	39.96	0.00
Myotis
Detector + Site	9	0.00	0.92	0.50
Detector	3	5.55	0.06
Site	8	8.67	0.01
Detector + Date	5	9.10	0.01
NULL	2	10.43	0.00
Date	4	13.83	0.00

Table 4. Estimates and 85% confidence intervals (CI) for independent beta regression models comparing proportion of calls identified to each frequency group by Anabat Swifts and AudioMoths. AudioMoth is used as the reference variable for Detector and Site 1 is the reference variable for Site.

Model	Fixed Effect	Estimate	85% CI
Proportion Low ~ Detector + Site	Intercept	1.255	0.807	1.702
	Detector (Swift)	−0.567	−0.916	−0.218
	Site (2)	−2.812	−3.446	−2.178
	Site (3)	−2.301	−2.896	−1.706
	Site (4)	−2.340	−2.938	−1.743
	Site (5)	−3.510	−4.205	−2.816
	Site (6)	−3.683	−4.391	−2.974
	Site (7)	−2.587	−3.202	−1.971
Proportion Mid ~ Site	Intercept	−1.468	−1.890	−1.046
	Site (2)	2.660	2.076	3.245
	Site (3)	2.237	1.673	2.800
	Site (4)	2.557	1.978	3.135
	Site (5)	3.536	2.884	4.188
	Site (6)	2.455	1.882	3.028
	Site (7)	2.792	2.200	3.385
Proportion Myotis ~ Detector + Site	Intercept	−2.595	−2.978	−2.213
	Detector (Swift)	0.685	0.420	0.949
	Site (2)	−0.097	−0.586	0.392
	Site (3)	0.081	−0.395	0.556
	Site (4)	−0.417	−0.933	0.099
	Site (5)	−0.630	−1.166	−0.094
	Site (6)	0.962	0.533	1.391
	Site (7)	−0.308	−0.814	0.199

Discussion

Our study tested how a less expensive passive acoustic device compared to a more expensive device in regards to data acquisition and quality of recorded high-frequency echolocation calls. We showed that the more expensive passive acoustic device recorded echolocation calls of higher quality than the less expensive device. However, when examining proportions of echolocation calls identified to phonic group, we found a discrepancy in recording rates between device types for the low-frequency and Myotis groups. Together, these results demonstrate that a less expensive device can be useful in collecting data for large-scale passive acoustic monitoring studies that require many sensors but may be inappropriate for studies that require fine details like species-level identification or focus on species vocalising in the high-frequency range.

There are not many published studies in the last two decades that compare passive acoustic devices. However, discrepancies in passive acoustic devices to record bat echolocation have been noted in previous work (Downes 1982; Forbes and Newhook 1990; Waters and Walsh 1994; Parsons 1996; Fenton et al. 2001; Adams et al. 2012; Kaiser and O’Keefe 2015). For example, some Anabat passive acoustic devices (models II and SD2, now retired) record fewer high-frequency calls, like those emitted by Myotis, and calls of lower quality (i.e., lower highest frequency, shorter bandwidth, and shorter duration) compared to other brands of acoustic devices (Pettersson D980 – Fenton 2000; QMC S200 – Fenton et al. 2001; Avisoft Bioacoustics UltraSoundGate, ecoObs Batcorder 2.0, Elekon AG Batlogger, Wildlife Acoustics SongMeter SM2Bat – Adams et al. 2012; Wildlife Acoustics SongMeter SM2Bat+ – Kaiser and O’Keefe 2015). There are several factors that can affect how a device detects sound in the environment. Directional microphones may record fewer echolocation calls than omnidirectional microphones because they do not sample behind the microphone (Sprong et al. 2012), but overall range may be better for directional microphones (Limpens and McCracken 2004). However, omnidirectional microphones recorded fewer low and midrange calls, but more Myotis calls, than a directional microphone in a study in eastern North America (Kaiser and O’Keefe 2015). The AudioMoth has a small microphone, but it is omnidirectional, unlike the microphone we used for the Swift. Perhaps the small microphone in an AudioMoth is less sensitive to high-frequency calls than larger omnidirectional microphones. Further, the disparity we observed could be explained by the fact that some microphones may be more sensitive to certain frequency ranges than others; e.g. Anabats are more sensitive to low-frequency (<30 kHz) sounds (Adams et al. 2012; Kaiser and O’Keefe 2015). Weatherproofing can also affect how PAM devices record sound (Britzke et al. 2010; Osborne et al. 2023). In our study, the Swift and AudioMoth devices were not weatherproofed in the same manner. Anabat microphones are weatherproofed with a thin mylar membrane with a gold coating; this membrane might transmit sound differently than a thin plastic membrane over the AudioMoth microphone, which is what we used in our study.

The more expensive device recorded a higher proportion of calls identified to the Myotis phonic group than the less expensive device, which could be due to the greater detection range of the more expensive microphone and device sensitivity. Both devices were set to record at their maximum sampling rate, but the Swift had a maximum of 500 kHz, and the AudioMoth maximum sampling rate was 384 kHz. Both devices should have been able to record any of the bats occurring in our study area, but there were discrepancies in recording rates by phonic group and the two devices were only similar for the mid-frequency phonic group. The fact that the less expensive device recorded a lower proportion of high-frequency files could be an issue if trying to identify different Myotis species. Myotis is the second largest genus of mammals with over 100 species on six continents (Simmons and Cirranello 2021). Many of these species are morphologically similar and have overlapping echolocation call structures (e.g. Murray et al. 2001; Russo and Jones 2002). High-quality recordings that retain the higher frequencies are required for identifying this often-cryptic genus. This consideration is especially important when designing surveys in North America, where it may be desirable to use acoustics to survey for Myotis and Perimyotis species endangered by white-nose syndrome (U.S. Fish and Wildlife Service 2007, 2022; Canadian Wildlife Service 2013).

We assert that high-quality recordings retaining higher frequency components and longer duration pulses are needed for parsing fine-scale differences among calls of bat species with overlapping echolocation characteristics. If a device records calls of shorter call durations, shorter bandwidths, and lower maximum frequencies, it clips parts of the echolocation call that could be important for identification. Depending on the species, automated identification programmes already have varying success in identifying a bat call to species (Rydell et al. 2017), and often echolocation calls are assigned to a phonic group due to inability to differentiate between species (e.g. Kaiser and O’Keefe 2015; Beilke et al. 2021). Identifying to phonic group is still very useful as it can give good information about the ecology of bats in terms of what types of bats are using an area (Beilke et al. 2021). In areas of high bat diversity, it is more likely that there will be multiple closely-related species with similar echolocation calls due to the structure of echolocation calls being related to a species’ phylogenetic history (see Russo et al. 2018). To improve the efficacy of low-cost sensors for identification purposes, it will be important to improve and fine-tune microphones.

Despite lower-quality recordings, we argue that less expensive devices are useful for certain questions. We found that the AudioMoth recorded a higher proportion of calls classified in the low frequency phonic group compared to the Swift. There are many bats that echolocate <35 kHz (Fenton and Bell 1981), and a less expensive device that adequately records low-frequency sounds would be good for monitoring those species, even with lesser data quality. For example, the Hawaiian hoary bat (Lasiurus cinereus semotus) and the Florida bonneted bat (Eumops floridanus) are endangered species that echolocate <30 kHz (U.S. Fish and Wildlife Service 1998, 2013; Belwood and Fullard 1984; Bailey et al. 2017). Open-source devices that are less expensive may be beneficial for large-scale projects focused on low-frequency echolocators. Other studies focused on general bat presence or activity where species identification is less important could also apply less expensive devices despite lower-quality recordings. Less expensive devices decrease project costs, making large-scale ecological studies more feasible, especially when high-frequency sounds are not expected. For example, AudioMoths can be used to monitor wolves (Barber-Meyer et al. 2020) and have also been used to develop automated frog call identification (Lapp et al. 2021). AudioMoths can be used to monitor bat activity at the entrances to caves as a less invasive approach than entering caves to monitor cave-dwelling bats when cost is a constraint (Revilla-Martín et al. 2021). However, in cases where the less expensive device is used in conjunction with another type of device (e.g. Froidevaux et al. 2021), care should be taken to note that there could be differences in detections between the two types of devices.

Open source devices like the AudioMoth are highly customisable and can be programmed to record sounds across many frequencies, which is useful in monitoring many different bat species (Hill et al. 2018). However, the AudioMoth is currently constrained to recording full-spectrum audio files, which will require a lot of storage space if used for long-term studies. Recording in zero-crossing, or frequency division, which retains only the dominant frequency heard (Parsons and Szewczak 2009), could save space and extend battery life for longer field studies. Since our field deployment study, there have been more refinements to AudioMoth hardware and firmware and, in the future, it might be possible to record in a different format, with onboard filtering, or for longer periods (Open Acoustic Devices 2020).

It is critical to develop more low-cost sensors and improve the quality of data that the devices can record to aid in large-scale conservation studies in a changing world. Open source devices like the AudioMoth or other types of sensors using software such as Arduino (www.arduino.cc) help advance the field of conservation by democratising access to monitoring devices for researchers around the globe (Hill et al. 2019). Because there are many threats to bats such as white-nose syndrome, habitat degradation, and wind energy (Fenton 1997; Blehert et al. 2009; Ellison 2012), and because bats can move long distances, being able to monitor their populations over large areas is critical in being able to conserve this group of animals.

Ethical statement

We followed institutional and national ethical guidelines for scientific research in the United States.

Acknowledgements

We thank Kathryn Bulliner and the Missouri Department of Conservation for procuring the AudioMoth devices. We thank technicians at Indiana State University for help in collecting the data.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

Data available from the Illinois Data Bank; doi.org/10.13012/B2IDB-4200947_V1.

Additional information

Funding

This work was funded by the Missouri Department of Conservation under Cooperative Agreement No. 418.