Acoustic target classification of zooplankton using machine learning

Student thesis: Doctoral ThesisDoctor of Philosophy (awarded by UHI)

Abstract

Acoustic target classification of zooplankton is challenging due to their small size, weak target strengths (TS), and tendency to form dense, mixed-species aggregations, which limits the effectiveness of conventional methods based on volume backscatter. Broadband echosounders measure backscatter over a continuous frequency range, providing more information for classification compared to narrowband. Through pulse-compression processing, they also have improved ability to resolve weakly backscattering, densely aggregated targets such as individual zooplankton. However, broadband echosounders are associated with increased data volume and complexity, necessitating the use of automated methods such as machine learning (ML). The feasibility of individual-scale classification of zooplankton using machine learning was evaluated through ex-situ experiments with commercially available broadband echosounders. TS-frequency spectra (‘target spectra’; 283-383 kHz) of copepods (Paraeuchaeta norvegica) and euphausiids (Thysanoessa raschii and Meganyctiphanes norvegica) were measured in a tank, and supervised ML classifiers were trained to differentiate them. The best-performing classifiers had a classification accuracy of 96%. However, obtaining labelled training data is a hurdle to field application. A novel method, using an unsupervised one-class support vector machine trained using only measurements of a species of interest (P. norvegica), was able to differentiate this species from other zooplankton in a community sample with a class-weighted F1 score of 0.71-0.85. An alternative method, using model-predicted target spectra as training data, was evaluated on measurements (185-255 kHz) of a mixed assemblage of Arctic mesozooplankton in a purpose-built cage. However, classifiers were unable to differentiate the mesozooplankton classes (copepods, euphausiids, chaetognaths, and hydrozoans) reliably due to the similarity of their modelled spectra (mean class-weighted F1 score: 0.68). Results suggest that individual-scale classification of zooplankton using machine learning is a viable approach for groups with similar material properties but discrete size distributions. Obtaining labelled training data remains a significant challenge, but cage experiments offer an effective solution.
Date of Award20 Oct 2023
Original languageEnglish
Awarding Institution
  • University of the Highlands and Islands
SponsorsNatural Environment Research Council (NERC), Norwegian Research Council & Marine Alliance for Science and Technology for Scotland (MASTS)
SupervisorKim Last (Supervisor) & Finlo Cottier (Supervisor)

Cite this

'