The use of automated identification of bat echolocation calls in acoustic monitoring: A cautionary note for a sound analysis
Introduction
Bats are one of the most diverse groups of living vertebrates (Tsang et al., 2015). Because of their sensitivity to human impact on ecosystems, bat populations are declining in many regions of the world (Hutson et al., 2001, Voigt and Kingston, 2016). For the same reason, bats may also serve as effective indicators of human impact on the biota (Jones et al., 2009, Russo and Jones, 2015). Therefore, there is an ever growing need of developing accurate and robust methods for surveying bat populations and monitoring their trends (Meyer, 2015).
Except for flying foxes (family Pteropodidae), all bats are laryngeal echolocators, i.e. they navigate in the dark and detect prey or food items by broadcasting (mostly ultrasonic) echolocation calls and perceiving the elaborate echoacoustic images of their surroundings (Jones and Holderied, 2007). They do so in a species-specific way according to the phylogeny, habitat use and biology of a given species (e.g. Arita and Fenton, 1997, Schnitzler et al., 2003, Skiba, 2003, Jones and Holderied, 2007). By means of bat detectors, echolocation has been widely exploited by researchers to reveal the presence of bats, investigate their behaviour, assess habitat use and estimate population sizes and trends. Bat detectors are devices which transform ultrasound into audible sound and, in the more recent versions, allow the storage of sound recordings for further processing and analysis. Similar approaches, yet without conversion of high frequency calls into the audible range, have also been used in birds (Scott, 2008) and frogs among others (Depraetere et al., 2012).
Bat detectors are highly valuable and thus irreplaceable tools for the study of bats. Because bats are nocturnal, elusive, hard to catch and often sensitive to disturbance, surveying their presence by bat detectors offers a non-invasive approach that complements effectively more invasive techniques such as roost surveys or captures (Flaquer et al., 2007). Species that roost in crevices and go unnoticed in roost surveys, as well as those that avoid nets because they fly too high up may still be recorded with bat detectors. In some cases, acoustic methods may outperform direct observation as they may help separate cryptic species whose morphology alone would be of little help (Barratt et al., 1997).
There are technical limitations to the number of species revealed by acoustic surveys, however, since species emitting low-amplitude calls, highly directional calls or high-frequency echolocators whose pulses undergo strong atmospheric attenuation, can be overlooked (e.g. Anderson and Racey, 1991). Although a combination of techniques is advisable to make sure all species occurring in a given area are detected (Flaquer et al., 2007), the technological progress characterizing bat detectors over the last 15 years ca. has made them increasingly important research tools. The most recent high-sampling rate models may record ultrasound in real-time and some are designed for long-term, unattended recording.
A long-standing, controversial issue regarding the application of bat detectors is how to use them to accomplish reliable acoustic identification of bat species. Unlike sound used by animals to communicate, which often shows a stereotyped structure, echolocation has primarily a sensorial function, so bat calls are very different from bird songs (Barclay, 1999). As it crosses different habitats, a bat may alter the structure of its echolocation calls dramatically to best match habitat structure: this results in an impressive intraspecific (in fact intra-individual) variation in call shape (e.g. Obrist, 1995, Fenton et al., 2004, Mora et al., 2005, Jones and Siemers, 2011). Conversely, echolocation call design often converges among species that use the same habitat or are adapted to similar habitat structure (Schnitzler and Kalko, 2001, Siemers et al., 2001). As a result of intraspecific variation and interspecific convergence in call design, call spectral and temporal characteristics widely overlap among species. Therefore, although calls from some species are easy to tell apart or even unmistakable, others are hard or impossible to recognize. The picture is complicated by further sources of variation such as atmospheric attenuation (Lawrence and Simmons, 1982), intraspecific geographic variation (Russo et al., 2007, Sun et al., 2013), sex (Russo et al., 2001, Puechmaille et al., 2014), age (Jones and Ransome, 1993), body condition (Puechmaille et al., 2014) or temporary changes in call structure induced by the presence of conspecifics to avoid sonar jamming (Ulanovsky et al., 2004).
Humans may classify bat calls in two main ways, qualitatively or quantitatively. When bat detectors appeared on the scene, a first approach adopted by bat workers was to rely on the personal, subjective assessment of the sound output by the detector to attempt species identification (Ahlén, 1981). At that time this was the only way to use bat detectors since the first devices used heterodyne technology, which degrades sound structure making sound analysis impossible. Noticeably, when time expansion became popular, offering a relatively cheap way of recording sound suitable to further quantitative analysis and use the latter approach to identify species, some aficionados of qualitative identification insisted on applying it to time-expanded sound (Ahlén and Baagøe, 1999), yet their approach did not get too far. The main argument against qualitative classification is that even if some people might reliably tell the various bat species apart by ear (achievable in regions where few species occur), their method would be hard to replicate being so subjective to remind “art” rather than “science”.
The popularity of quantitative classification, based on the measurement of spectral and temporal features of bat calls, has increased with the advent of time-expansion detectors and easy-to-use sound analysis software, and more recently with the launching of devices on the market that carry out real-time ultrasound recording. Although measuring bat pulse features to attempt identification is a more objective approach, it still does not solve the problem of differentiating acoustically between all species of a local ensemble based on their echolocation calls alone. Because multiple variables can be measured out of a bat call, the problem of species discrimination can be best tackled with a multivariate approach.
Researchers have dealt with developing automated methods of bat call classification since the end of 1990s, when the first papers describing applications of multivariate discriminant function analysis or neural network were published (Zingg, 1990, Vaughan et al., 1997a, Parsons and Jones, 2000, Russo and Jones, 2002). Since then, a wealth of new methods have been applied to try to solve the problem, such as, to mention some, algorithms of pattern recognition (Obrist et al., 2004), support vector machines (Redgwell et al., 2009), hierarchical ensembles of neural networks (Redgwell et al., 2009, Walters et al., 2012), geometric morphometry (MacLeod et al., 2013), machine learning (Skowronski and Harris, 2006), CART (Preatoni et al., 2005) and random forest classification (Marckmann and Runkel, 2010). In all such cases it is an operator-independent classifier to evaluate the structure of an unknown call and attribute the latter to a species. Therefore, the main advantage offered by automated classifiers is that they do not require the intervention of a human operator in the classification phase, which circumvents the pitfalls of subjectivity in the process (e.g. Parsons and Jones, 2000). Jennings et al. (2008) carried out a comparison of the performances of human vs. non-human classifiers (artificial neural networks) and found that the latter outperformed 75% of the humans. It was also found that for humans additional experience beyond a first period of ca. 1 year brings about little improvement in identification skills (Jennings et al., 2008).
The detailed functioning of such classifiers, as well as their pros and cons, are out of the scope of this paper, but broadly speaking the underlying principle is the comparison of an unknown call we wish to identify with a library of calls of known identity: the unknown call is then attributed to the species whose reference calls are structurally closest. All automated classifiers work out a probability value of correct classification which the operator may use to evaluate the reliability of the response.
It is worth remarking that the quality of call library affects crucially the quality of responses (Clement et al., 2014). A first fundamental issue is that the origin of reference calls must be known with certainty, i.e. they must have been recorded from bats whose identity was established through other methods. Such bats may have been identified from morphology and recorded on roost emergence or hand release; in some cases echolocation calls may have been extracted from sequences of free-flying bats including species-specific social calls used for unambiguous classification.
A second important aspect is that the reference library should cover as much as possible the (often overwhelming) intraspecific call variation related to habitat structure, geographical differences, etc. Capturing such variation is objectively a most challenging task (Russo and Jones, 2002, Clement et al., 2014). Recording conditions may also affect call structure: for example, calls broadcast by hand-released bats often exhibit higher frequency values as well as shorter durations and interpulse intervals than typical search-phase calls.
The first articles presenting automated classification of bat calls had the main aim of testing whether different classifiers could improve classification rates but the resulting “prototype” algorithms were not meant to be circulated as the solution to the identification problem (Vaughan et al., 1997a, Parsons and Jones, 2000, Russo and Jones, 2002). For this reason, although the first automated classifiers were applied to specific case studies, such as the assessment of habitat use in specific areas (Vaughan et al., 1997b, Russo and Jones, 2002), they were not made publicly available as the authors were conscious of their limitations and inherent risks if uncritically adopted (effects of reference datasets or potential biases arising when classifiers are used outside the geographic region they had been devised for, etc.). Since the first pioneering experiments, more recent work has mainly brought about further improvements in the classification performances.
Walters et al. (2012) have made a considerable step forward by developing a classifier from 1350 reference calls for 34 European species based on a hierarchy of ensembles of artificial neural networks. This freely available classifier obtained high performances in identifying several species and has been proposed as a tool for the implementation of a global bat bioindicator.
Over the following years, the production of automated classifiers escalated rapidly: some are now available on purchase, others are free but only usable in association with commercially sold software or hardware. Such products have filled a vacant niche on the market mostly in relation to the expanding monitoring efforts related to the development of wind power production worldwide. Following concern over bat fatalities at wind turbines (e.g. Rydell et al., 2010, Lehnert et al., 2014), governments of many countries, e.g. those of the European Union, have made it mandatory to carry out pre-construction assessments of bat activity or the monitoring of existing plants. Acoustic surveys of bats are the main approach used by consultancies carrying out such work. This typically implies going through thousands of recordings made by unattended units that are triggered automatically by bat calls (and, not rarely, by unwanted noise), which makes the manual selection and identification of calls extremely time-consuming besides requiring strong identification skills. Clearly, automated classifiers have been greeted as the solution to such issues. This is a turning point in the story we have told so far, with its own strengths and weaknesses.
Section snippets
The automated processing of bat pass recordings
Offering a comparison of the various alternatives present on the market is out of our scope, which is instead to discuss how the progressive spread of automated classification among bat workers will likely influence key aspects of bat studies and its potential implications for conservation. We therefore deliberately refrain from listing the several automated software packages available to date and discussing their features – this information is easily obtainable by searching the web.
Needless to
Validating software performances
Clement et al. (2014) examined an important aspect of automatic classification, namely the effects of filters and validation datasets on classification performances. In the first automated classification experiments (Vaughan et al., 1997a, Russo and Jones, 2002) spectral and temporal measurements were taken by hand. Although in these cases the operator is fully aware of what is being measured, the process is time-consuming and somewhat subjective. More recent automated classifiers, including
Conclusions and implications for conservation
Wrong information is worse than no information: biases in the assessment of bat distribution or habitat preferences may lead to wrong management decisions with serious consequences for the conservation of bat populations. We remark that the problem of bat call identification should be regarded as a serious practical issue rather than a mere academic exercise.
Acoustic surveys may tackle different objectives and the potential impact of automated classifiers on data quality will differ
Acknowledgments
We are grateful to two anonymous reviewers for their valuable comments on a first manuscript version. Thanks also go to Gareth Jones and Jens Rydell for useful discussion.
References (52)
- et al.
Flight and echolocation in the ecology and evolution of bats
Trends Ecol. Evol.
(1997) - et al.
Monitoring animal diversity using acoustic indices: implementation in a temperate woodland
Ecol. Indic.
(2012) Methodological challenges in monitoring bat population- and assemblage-level changes for anthropogenic impact assessment
Mamm. Biol.
(2015)- et al.
Bats as indicators: an introduction
Mamm. Biol.
(2015) - et al.
From spatial orientation to food acquisition in echolocating bats
Trends Ecol. Evol.
(2003) - et al.
Poor effectiveness of Natura 2000 beech forests in protecting forest-dwelling bats
J. Nat. Conserv.
(2015) Field identification of bats and survey methods based on sounds
Myotis
(1981)- et al.
Use of ultrasound detectors for bat studies in Europe: experiences from field identification, surveys, and monitoring
Acta Chiropterol.
(1999) - et al.
Feeding behaviour of captive brown long-eared bats Plecotus auritus
Anim. Behav.
(1991) Bats are not birds – a cautionary note on using echolocation calls to identify bats: a comment
J. Mammal.
(1999)