Home // COCE 2025, The Second International Conference on Technologies for Marine and Coastal Ecosystems // View article


Machine Learning Using Physics-Based Backscattering Modeling for Acoustic Identification of Mesopelagic Organisms

Authors:
Sander Andre Berg Marx
Babak Khodabandeloo
Ketil Malde

Keywords: Acoustic target classification; machine learning; Convolutional neural network; prolate spheroid backscattering modeling.

Abstract:
Target Strength (TS) is the logarithmic measure of the backscattered acoustic energy reflected toward a sound source when an acoustic wave encounters an organism. It depends on the organism’s size, shape, material properties, orientation, and the frequency of the wave, and thus carries information useful for identifying and characterizing marine organisms. Advanced broadband echosounders now allow detailed TS measurements of marine organisms over nearly continuous frequency ranges, which provide valuable information for biomass estimation and ecosystem monitoring. However, interpreting these TS measurements has traditionally relied on manual classification, which makes it difficult to extract biological characteristics for target classification or ecological analysis, especially given the complexity of broadband data. Physics-based backscattering models are versatile tools for modeling the TS frequency response given the shape and material properties of the scatterers. In this study, we employ an exact prolate spheroid model, representative of many marine organisms, to simulate broadband TS spectra for training machine learning models. These models aim to classify and characterize targets based on their TS frequency signatures. A hybrid one-Dimensional Convolutional Neural Network (1D-CNN) is proposed for simultaneous classification (gas- vs. liquid-filled) and regression of geometric properties and compared against KNN, SVM, and RF. Results show that while all models achieved perfect classification accuracy, the hybrid 1D-CNN clearly outperformed the others in parameter estimation. This demonstrates that simulation-driven machine learning can help overcome data scarcity and enable automated acoustic identification of mesopelagic organisms.

Pages: 32 to 37

Copyright: Copyright (c) IARIA, 2025

Publication date: October 26, 2025

Published in: conference

ISBN: 978-1-68558-329-3

Location: Barcelona, Spain

Dates: from October 26, 2025 to October 30, 2025