Comparing automated species identification classifiers for acoustic bat data from Anabat and AudioMoth detectors

Perks, Samantha J ORCID: https://orcid.org/0000-0003-1893-8059, O'Connell, Mark ORCID: https://orcid.org/0000-0003-3402-8880 and Goodenough, Anne E ORCID: https://orcid.org/0000-0002-7662-6670 (2026) Comparing automated species identification classifiers for acoustic bat data from Anabat and AudioMoth detectors. Acta Chiropterologica. (In Press)

Text (Peer-reviewed version) 16387 Perks (2026) Comparing automated species identification classifiers.pdf - Accepted Version Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial No Derivatives 4.0. Download (1MB)

Abstract

Passive Acoustic Monitoring (PAM) is becoming an increasingly popular method for surveying and monitoring bats within conservation and scientific research contexts, as well as for legislative complacence. Although PAM protocols have the potential to be standardisable and scalable, they typically produce vast acoustic datasets that require considerable time and ecological skills to analyse manually. With the continually evolving capabilities of Artificial Intelligence (AI), several automated classifiers now exist to classify bat guilds, which have the potential to streamline analysis workflows considerably. However, uncertainty remains regarding the reliability of such classifiers, particularly as regards how their outputs differ from one another and how their performance is impacted by differential recording quality from different types of detector. Here, agreement between three commonly used classifiers (Kaleidoscope Pro, BatClassify in Anabat Insight, and the British Trust for Ornithology Acoustic Pipeline) is investigated for acoustic recordings from commercial bat detectors (Anabat Swift; n = 13,582 recordings) and open-source acoustic recorders (AudioMoth; n = 30,574 recordings). Habitat type (riparian, woodland, wood pasture, arable) and bat species both affected the level of agreement between detectors. Disagreement most prevalent on recordings from more structurally complex habitats such as woodland and where a species within the genus Myotis was classified by at least one classifier. There were also differential interactions between habitat and taxonomy depending on detector type. For Anabat recordings, disagreement was high for Barbastella and Plecotus in woodland and wood pasture relative to riparian and arable; for AudioMoth recordings disagreement was high for Nyctalus/Eptesicus in woodland and Rhinolophus in riparian habitat. These findings suggest that classifiers must be used with caution and in conjunction with manual auditing carried out by suitable skilled practitioners, especially when errors in classification have consequences within formal legislative frameworks.

Item Type:	Article
Article Type:	Article
Uncontrolled Keywords:	PAM; Bats; Automated detection; Anabat; AudioMoth; Bioacoustics; Automated classification; Automated identification
Subjects:	Q Science > QL Zoology > QL605 Chordates. Vertebrates > QL737.35 Chiroptera
Divisions:	Schools and Research Institutes > School of Education, Health and Sciences
Depositing User:	Rhiannon Goodland
Date Deposited:	29 Jun 2026 11:05
Last Modified:	29 Jun 2026 11:15
URI:	https://eprints.glos.ac.uk/id/eprint/16387

University Staff: Request a correction | Repository Editors: Update this record