Browse by author
Lookup NU author(s): Dr Matthew Sharpe
This work is licensed under a Creative Commons Attribution 4.0 International License (CC BY 4.0).
© 2026 The Author(s). Marine Mammal Science published by Wiley Periodicals LLC on behalf of Society for Marine Mammalogy.Large volumes of passive acoustic data are collected by researchers, governments, and conservationists across the globe to monitor species for population assessments and conservation objectives. An analysis bottleneck often exists due to the lack of tools to process large datasets and identify detected signals to the species level. Deep learning has been shown to enable fast and effective detection of species-specific vocalizations. This study developed a Risso's dolphin (Grampus griseus) binary echolocation click classifier using a Long Short-Term Memory neural network. This classifier was trained and validated on 872 visually verified single species encounters collected during towed hydrophone surveys, primarily in the Northeast Atlantic. The classifier was subsequently tested on an independent set of 31 h of visually verified data collected from static and drifting recorders in Scottish and English waters, comprising four different delphinid species (including Risso's dolphins) as well as various sources of anthropogenic noise. The best classifier achieved an F1 score of 0.98 (precision 0.96, recall 0.98). While care must be taken when applying deep learning classifiers to new datasets, the performance of this model on independent data highlights its ability to generalize well within the study region, even when using data collected on different recording platforms.
Author(s): Webber T, Risch D, Hastie G, Sinclair R, Beck S, Berggren P, Boisseau O, Dyke K, Hartny-Mills L, Lacey C, Quer S, Sharpe M, Walters A, Young KF, van Geel N
Publication type: Article
Publication status: Published
Journal: Marine Mammal Science
Year: 2026
Volume: 42
Issue: 2
Print publication date: 20/04/2026
Online publication date: 20/04/2026
Acceptance date: 13/04/2026
Date deposited: 06/05/2026
ISSN (print): 0824-0469
ISSN (electronic): 1748-7692
Publisher: John Wiley and Sons Inc
URL: https://doi.org/10.1111/mms.70180
DOI: 10.1111/mms.70180
Data Access Statement: The best performing classifier developed here is openly accessible. The models can be run on both CPUs and CUDA- enabled GPUs, either in dependently or directly within PAMGuard. Instructions for use and re training are available here: github.com/tomwebber96/EcholocationClassification. The classifier is also provided in a Hierarchical Data Format (HDF5) which allows additional training data to be added to the existing model if required. The model architecture can also be easily extracted should a user wish to train the model from scratch and/or change the model architecture to increase performance in a new use case.
Altmetrics provided by Altmetric
