2025_programme: Power and accuracy trade-offs for machine learning methods applied to detection of underwater sound sources

Back


  • Day: June 17, Tuesday
      Location / Time: D. CHLOE at 14:50-15:10
  • Last minutes changes: -
  • Session: 06. Enhancing underwater acoustic sensing through machine learning
    Organiser(s): Cheng Chi
    Chairperson(s): Peng Xiao
  • Lecture: Power and accuracy trade-offs for machine learning methods applied to detection of underwater sound sources
    Paper ID: 2301
    Author(s): William Butler, Harrison Smith, Marios Impraimakis, Andrew Barnes, Alan Hunter
    Presenter: William Butler
    Abstract: Passive Acoustic Monitoring (PAM) is a crucial tool for the non-invasive detection, observation and population estimation of marine mammals, the determination and tracking of wildlife diversity as well as the detection and tracking of other underwater noise-emitting objects. Convolutional Neural Networks (CNNs) and other types of Artificial Neural Networks (ANNs) have been used to classify acoustic data. However, ANNs can require significant computational workload at inference and are therefore power intensive, while PAM is often employed on long duration, low-powered autonomous devices due to its low power requirements. In this paper, various ANN based methodologies for classification of seal vocalisations from PAM data are compared on both classification validity and runtime computational cost at inference. In addition, the power required to perform necessary preprocessing steps will be analysed; in particular, the use of FFTs to decompose audio data into spectrograms for use with CNN based image classification techniques will be compared against a 1D kernel CNN with a similar parameter space, trained on the raw audio input. The training behaviour, overall performance and class specific metrics of 1D-kernel CNN and 2D-kernel CNN are analysed in the context of their full system power requirements, including the pre-processing steps.
      Download the full paper
    This paper is a candidate for the "Prof. Leif Bjørnø Best Student Paper Award (for students under 35)"
  • Corresponding author: Mr William Butler
    Affiliation: University of Bath
    Country: United Kingdom

Back