Wavelet Scattering Transforms Applied to Whale Vocalisations

In this dissertation by publication, presented as three journal papers, we investigate the application of wavelet scattering transforms to whale vocalisations. We first consider how wavelet transforms can provide an alternative time-frequency decomposition for use in signal detectors. Detectors typically utilise the short-time Fourier transform (STFT), for which the continuous wavelet transform (CWT) serves as a suitable alternative. As an additional contribution, we improve upon the spectral entropy (SE) detector, which is known to be superior in terms of signal detection compared to energy detectors. In the second publication, we expand on the connection of the CWT to wavelet scattering. Wavelet scattering is a feature extraction method akin to the filters of a convolutional neural network frontend and mel-frequency cepstral coefficients. We show that wavelet scattering coefficients can be combined with a SE detector, which can be fed in to a classification system. We further improve upon the SE detector and provide a critical discussion on the challenges faced on a real-world dataset. Finally, the concepts established in the previous papers are used to construct an entirely new and generalised form of the wavelet scattering transform — joint n-dimensional scattering. We extend the definition of one-dimensional scattering to an arbitrary number of signal dimensions by utilising separable filters. Joint scattering benefits from various computational and conceptual advantages, which have not been generalised to this extent in the current literature.

Grantor dc:publisher

Stellenbosch : Stellenbosch University

Year dc:date.issued

2026

Author dc:creator

• Rademan, Marco Wiehann

Advisors dc:contributor.advisor

• Versfeld, Daniel J. J.
• Du Preez, Johan A.

Language dc:language.iso

en

Repository record dc:identifier.uri

https://scholar.sun.ac.za/handle/10019.1/135919

OAI identifier oai:identifier

oai:scholar.sun.ac.za:10019.1/135919