Unravelling the Marine Symphony: An Automated Approach to Identifying Delphinidae Signature Whistles through Passive Acoustic Monitoring

Abstract

As global climates change and coastal urbanisation intensifies, assessing the ecological impact on marine environments becomes crucial. Rising sea temperatures and habitat degradation pose significant challenges. Monitoring top predator populations, like delphinidae in ecosystems like the North East of England, provides insights into marine ecosystem health and aids in developing effective conservation strategies. By studying delphinidae populations, researchers can gain a better understanding of the health of the marine food chain, identify changes in ecosystem structure, and develop more effective conservation strategies. Passive Acoustic Monitoring (PAM) is a non-invasive method to monitor delphinidae by deploying acoustic recorders in static locations for extended periods. Analysing the accumulated acoustic data requires significant time and effort. PAM can be used to monitor bottlenose dolphin abundance as the species produces signature whistles (a distinct whistle vocalisation that every dolphin produces and is unique to themselves). These types of vocalisations enable the creation of a catalogue to identify individuals within a pod and subsequently calculate abundance estimates to monitor population dynamics using a capture-recapture methodology. This thesis presents a framework for automatically identifying delphinidae that produce signature whistles. It begins with describing the data collection methods used to capture bottlenose dolphin vocalisations along the North East of England coastline. This primary data collection effort is crucial to the thesis as all remaining work is built upon it. A discussion on employing citizen science to aid initial whistle extractions from the collected data presented. Next, WaveGAN (a popular audio generative adversarial network) is used to generate synthetic dolphin whistles, which leads to an investigation of the effects of signal-to-noise ratio on the production of synthetic data. Using the initial and synthetic data, a whistle detector is developed using convolutional neural networks. Next, an investigation of unsupervised and supervised learning methods to develop feature embeddings for a signature whistle classification model is undertaken using a signature whistle dataset. A triplet network is decided upon. Finally, data augmentation is investigated to enhance the signature whistle classification model achieving an overall accuracy of 82.1% and a weighted F1 score of 0.819. Automating this time-consuming task is expected to enable researchers to focus on data analysis and interpretation, providing valuable insights into the ecological health of marine environments amid the challenges posed by climate change and coastal development.