Impact of transfer learning methods and dataset characteristics on generalization in birdsong classification DOI Creative Commons
Burooj Ghani, Vincent J. Kalkman, Robert Planqué

и другие.

Scientific Reports, Год журнала: 2025, Номер 15(1)

Опубликована: Май 9, 2025

Abstract Animal sounds can be recognised automatically by machine learning, and this has an important role to play in biodiversity monitoring. Yet despite increasingly impressive capabilities, bioacoustic species classifiers still exhibit imbalanced performance across habitats, especially complex soundscapes. In study, we explore the effectiveness of transfer learning large-scale bird sound classification various conditions, including single- multi-label scenarios, different model architectures such as CNNs Transformers. Our experiments demonstrate that both finetuning knowledge distillation yield strong performance, with cross-distillation proving particularly effective improving in-domain on Xeno-canto data. However, when generalizing soundscapes, shallow exhibits superior compared distillation, highlighting its robustness constrained nature. study further investigates how use multi-species labels, cases where these are present but incomplete. We advocate for more comprehensive labeling practices within animal community, annotating background providing temporal details, enhance training robust classifiers. These findings provide insights into optimal reuse pretrained models advancing automatic recognition.

Язык: Английский

AudioProtoPNet: An interpretable deep learning model for bird sound classification DOI Creative Commons
René Heinrich, Lukas Rauch, Bernhard Sick

и другие.

Ecological Informatics, Год журнала: 2025, Номер unknown, С. 103081 - 103081

Опубликована: Фев. 1, 2025

Язык: Английский

Процитировано

2

All thresholds barred: direct estimation of call density in bioacoustic data DOI Creative Commons
Amanda K. Navine,

Tom Denton,

Matthew J. Weldy

и другие.

Frontiers in Bird Science, Год журнала: 2024, Номер 3

Опубликована: Апрель 24, 2024

Passive acoustic monitoring (PAM) studies generate thousands of hours audio, which may be used to monitor specific animal populations, conduct broad biodiversity surveys, detect threats such as poachers, and more. Machine learning classifiers for species identification are increasingly being process the vast amount audio generated by bioacoustic expediting analysis increasing utility PAM a management tool. In common practice, threshold is applied classifier output scores, scores above aggregated into detection count. The choice produces biased counts vocalizations, subject false positive/negative rates that vary across subsets dataset. this work, we advocate directly estimating call density : proportion windows containing target vocalization, regardless score. We propose validation scheme in body data obtain, through Bayesian reasoning, probability distributions confidence both positive negative classes. use these predict site-level densities, distribution shifts (when defining characteristics change). These methods outputs any binary operating on fixed-size input windows. test our proposed real-world study Hawaiian birds provide simulation results leveraging existing fully annotated datasets, demonstrating robustness variations model quality.

Язык: Английский

Процитировано

8

Counting the chorus: A bioacoustic indicator of population density DOI Creative Commons
Amanda K. Navine, Richard J. Camp, Matthew J. Weldy

и другие.

Ecological Indicators, Год журнала: 2024, Номер 169, С. 112930 - 112930

Опубликована: Дек. 1, 2024

Язык: Английский

Процитировано

3

HawkEars: A regional, high-performance avian acoustic classifier DOI Creative Commons

Jan Huus,

Kevin G. Kelly,

Erin M. Bayne

и другие.

Ecological Informatics, Год журнала: 2025, Номер unknown, С. 103122 - 103122

Опубликована: Март 1, 2025

Язык: Английский

Процитировано

0

Impact of transfer learning methods and dataset characteristics on generalization in birdsong classification DOI Creative Commons
Burooj Ghani, Vincent J. Kalkman, Robert Planqué

и другие.

Scientific Reports, Год журнала: 2025, Номер 15(1)

Опубликована: Май 9, 2025

Abstract Animal sounds can be recognised automatically by machine learning, and this has an important role to play in biodiversity monitoring. Yet despite increasingly impressive capabilities, bioacoustic species classifiers still exhibit imbalanced performance across habitats, especially complex soundscapes. In study, we explore the effectiveness of transfer learning large-scale bird sound classification various conditions, including single- multi-label scenarios, different model architectures such as CNNs Transformers. Our experiments demonstrate that both finetuning knowledge distillation yield strong performance, with cross-distillation proving particularly effective improving in-domain on Xeno-canto data. However, when generalizing soundscapes, shallow exhibits superior compared distillation, highlighting its robustness constrained nature. study further investigates how use multi-species labels, cases where these are present but incomplete. We advocate for more comprehensive labeling practices within animal community, annotating background providing temporal details, enhance training robust classifiers. These findings provide insights into optimal reuse pretrained models advancing automatic recognition.

Язык: Английский

Процитировано

0