2025_programme: Machine/Deep Learning categorisation of sub-kilohertz Arctic soundscapes

• Day: June 19, Thursday
    Location / Time: B. ERATO at 17:20-17:40
• Last minutes changes: -
• Session: 01. Acoustics in Ocean Observation Systems
  Organiser(s): Jaroslaw Tegowski, Philippe Blondel, Hanne Sagen
  Chairperson(s): Jaroslaw Tegowski, Philippe Blondel
• Lecture: Machine/Deep Learning categorisation of sub-kilohertz Arctic soundscapes [Invited]
  Paper ID: 2276
  Author(s): Jonathan Cleverly, Philippe Blondel, Hanne Sagen, Espen Storheim, Matthew Dzieciuch
  Presenter: Jonathan Cleverly
  Abstract: Arctic soundscapes are being modified by climate change, which is greatly amplified in the region. With declining sea ice, cryophony (sounds of sea ice dynamic processes) is more variable, while the presence of marine mammal vocalisations and anthropogenic noise are changing with time and space. These markers of the state of the Arctic Ocean are monitored using passive acoustic technologies, however there are still no standard practices for exploring soundscapes in this region. Here we investigate statistical and Machine/Deep Learning (ML/DL) approaches for categorising recordings from deep-water Arctic regions by their contents. Recordings were taken from long-term hydrophone deployments during projects: “Acoustic Ocean Under Melting Ice” (UNDER-ICE, Fram Strait, 2014-2016) and “Coordinated Arctic Acoustic Thermometry Experiment” (CAATEX, Nansen Basin, 2019-2020). To minimise power requirements, UNDER-ICE recorded for 130 s every 3 hours at a sampling rate of 1,953 Hz, whereas CAATEX recorded for 45 minutes every 12 hours at a sampling rate of 976 Hz. First, we explore the relationships between commonly used acoustic indices (acoustic complexity index, biodiversity index etc.) and Principal Component Analyses to obtain a coarse description of soundscape evolution. Then, we utilise unsupervised/self-supervised algorithms such as AVES (Hagiwara, 2022) or Ketos (MERIDIAN, 2020) to deduce the contents within spectrogram representations of recordings (e.g. whale calls, seismic sounds). Training datasets for ML/DL algorithms usually consider a broader frequency range (beyond 20 kHz), but it is not always feasible to use these higher sampling rates, thus the robustness of algorithms require testing with lower sample rate data, where the frequency content of the sounds is not always fully recorded. These techniques will be crucial for avoiding current bottlenecks in data processing, in particular in the Arctic, enabling more in-depth studies for marine mammal conservation and industrial regulation.
    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 Jonathan Cleverly
  Affiliation: University of Bath
  Country: United Kingdom