Back to search

University of Southampton

Marine soundscapes, acoustic indices and machine learning

Abstract

dc:description.abstract

Passive Acoustic Monitoring (PAM) provides complimentary data to other marine monitoring methods. Long-term data sets are now increasingly common, with information regarding the biophony (biotic noise), anthrophony (anthropogenic noise) and geophony (abiotic noise) used to inform on marine management questions and policies, monitor species distributions and abundance, and inform on habitat health or restoration progress. PAM data were broadly analysed at the soundscape scale, typically using direct measures of pressure and power, or at the sound source level through detection and classification, increasingly using deep learning. Acoustic Indices are signal processing metrics which have been developed primarily in terrestrial acoustics and can be applied to soundscape monitoring and as features for machine learning. This thesis explores the use of acoustic indices in shallow, highly variable soundscapes to provide further insight at the soundscape level and as the basis for a low temporal resolution classifier. To improve the interpretation of results and initial hyperparameter choices of acoustic indices, environmental and labelled data sets are used with computationally efficient modelling techniques to establish the drivers of indices over 8 moorings and 4 months of data from a hydrophone array deployed offshore western Scotland and Northern Ireland. General Additive Models are applied with modelled wind and tidal data as explanatory variables. Results show a high degree of variability in the sensitivity of indices across the months and locations. Comparative analyses between vessel and non-vessel data sets show all indices are driven by vessel noise, but to varying extents and with high regional, but lesser temporal, variations. Uniform Manifold Approximation and Projection was applied to multi-labelled data, showing clusters were driven by distinct sonotypes (measured features resulting from all sound sources present) rather than individual label categories or types. This shows environmental data sets can be used to provide insight into drivers of acoustic indices, and results provided evidence of the complexity of the underlying soundscapes. Motivated by unsupervised clustering of coherent sonotypes, a classifier using acoustic indices as features was applied to 4 categories of Delphinid Vocalisations, Sonar, Vessels and Ambient noise (the absence of vessels), across two years and locations from the same PAM data set. Pipelines were developed to use the same features for each task, thereby it was not reliant on user expertise in feature selection or model hyperparameter tuning. Classifiers were applied at a 20-minute resolution. The best classification results used data from all locations for vessel detection, with an area under the precision recall curve (PRAUC) of 0.63, and on sonar using data from a single location with a PRAUC of 0.76. Other classifiers tested with different numbers of features and on data from all locations or single locations did not perform well, with overfitting observed during training. Whilst promising, this classification approach requires further refinement to be practically applicable. Finally, data collected during the PhD from UK Seagrass meadows were analysed to establish the utility of PAM in this critical habitat. Spectral analyses derived from Sound Pressure Levels (SPLs) in the frequency range 0.01 – 1 kHz and direct depth measures show a tidal component in both seagrass meadows, but that tidal noise was not dominant. Diel trends in SPLs and in the Acoustic Complexity Index (ACI) for 5 – 20 kHz in both meadows show soundscapes are driven in part by biophony, and that the ACI could be used as a complimentary measure in future work.

Degree

thesis:*
Name dc:type.qualificationname
Ph.D.
Level dc:type.qualificationlevel
doctoral
Grantor dc:publisher.institution
University of Southampton
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Dell, Benedict
Advisors dc:contributor.advisor
  • White, Paul
  • Bull, Jonathan
  • Risch, Denise
ORCID person:identifier.orcid

Rights

Language dc:language
English, eng

Identifiers

dc:identifier.*

Chain of custody

source
Harvested from
University of Southampton
Base URL
eprints.soton.ac.uk/cgi/oai2
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Dell, Benedict. Marine soundscapes, acoustic indices and machine learning. doctoral thesis, University of Southampton, 2026. https://doi.org/10.5258/SOTON/01/D087