Robust detection of North Atlantic right whales using deep learning methods

Vickers, William (2021) Robust detection of North Atlantic right whales using deep learning methods. Doctoral thesis, University of East Anglia.

[thumbnail of 2022VickersWPhD.pdf]
Download (37MB) | Preview


This thesis begins by assessing the current state of marine mammal detection, specifically investigating currently used detection platforms and approaches of detection. The recent development of autonomous platforms provides a necessity for automated processing of hydrophone recordings and suitable methods to detect marine mammals from their acoustic vocalisations. Although passive acoustic monitoring is not a novel topic, the detection of marine mammals from their vocalisations using machine learning is still in its infancy. Specifically, detection of the highly endangered North Atlantic right whale (Eubalaena glacialis) is investigated. A large variety of machine learning algorithms are developed and applied to the detection of North Atlantic right whale (NARW) vocalisations with a comparison of methods presented to discover which provides the highest detection accuracy. Convolutional neural networks are found to outperform other machine learning methods and provide the highest detection accuracy when given spectrograms of acoustic recordings for detection.

Next, tests investigate the use of both audio and image based enhancements method for improving detection accuracy in noisy conditions. Log spectrogram features and log histogram equalisation features both achieve comparable detection accuracy when tested in clean (noise-free), and noisy conditions.

Further work provides an investigation into deep learning denoising approaches, applying both denoising autoencoders and denoising convolutional neural networks to noisy NARW vocalisations. After initial parameter and architecture testing, a full evaluation of tests is presented to compare the denoising autoencoder and denoising convolutional neural network. Additional tests also provide a range of simulated real-world noise conditions with a variety of signal-to-noise ratios (SNRs) for evaluating denoising performance in multiple scenarios. Analysis of results found the denoising autoencoder (DAE) to outperform other methods and had increased accuracy in all conditions when testing on an underlying classifier that has been retrained on the vestigial denoised signal. Tests to evaluate the benefit of augmenting training data were carried out and discovered that augmenting training data for both the denoising autoencoder and convolutional neural network, improved performance and increased detection accuracy for a range of noise types.

Furthermore, evaluation using a naturally noisy condition saw an increase in detection accuracy when using a denoising autoencoder, with augmented training and convolutional neural network classifier. This configuration was also timed and deemed capable of running multiple times faster than real-time and likely suitable for deployment on-board an autonomous system.

Item Type: Thesis (Doctoral)
Faculty \ School: Faculty of Science > School of Computing Sciences
Depositing User: Chris White
Date Deposited: 23 Aug 2022 09:15
Last Modified: 23 Aug 2022 09:15

Actions (login required)

View Item View Item