Open Collections

Lu, Xi

University of British Columbia

In the underwater realm, marine mammals rely heavily on acoustic signals for their communications. Being able to accurately and easily detect these signals can aid in studying factors such as creature presence and migratory habits. Additionally, recent research in the field of covert Underwater Acoustic Communications (UWAC) has begun to incorporate marine mammal signals to utilise naturally-occurring sounds. Compared to traditional methods, marine mammal signals allow transmissions to occur at higher power levels, increasing the range of covert communications. Commonly, there are two categories of acoustic marine mammal signals: clicks and whistles. Both can be detected by converting audio waveforms to spectrogram images, and the detection process is often done visually by human experts. This image data type is particularly important for whistle detection, as these tend to be lower power and have a narrow-band, time-varying frequency profile. Given the potential uses of these signals, the creation of accurate, consistent automated detection methods has been an active area of research. This thesis investigates the utility of Neural Networks (NNS) in application to dolphin whistle detection and generation. We seek to provide a detection pipeline which is robust to changing environments and requires no context-specific work to be done. This is accomplished by performing minimal preprocessing on data and utilising transfer learning from a large dataset into a newer, smaller one. Using these techniques, we are able to achieve detection accuracy greater than 95% for our tested models. For whistle generation, we investigate two methods known as Generative Adversarial Networks (GANS) and Denoising Diffusion Probabilistic Models (DDPMS), the latter of which is found to be more effective. We separate the task of generating synthetic realistic whistles into two steps: contour and variations. The end result is a cascaded DDPM system which generates whistles following these two steps. We demonstrate an iterative detection application to assess the efficacy of this generative method, integrating our synthetic samples into the task of improving automated signal detection.

Text

2024-01-05

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/87136

Electrical and Computer Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/