Advancing passive acoustic monitoring of tropical coral reefs with deep learning

Passive acoustic monitoring (PAM) can be used to deliver ecological insights and guide conservation strategies. However, the pace at which PAM can be gathered greatly exceeds the capacity of manual analysis to process these data. To maximise the insights that can be generated from PAM, reliable automated analysis is needed. Deep learning (DL) represents a promising technology for automated analysis of PAM data but remains largely underexplored for tropical coral reefs. In this thesis I investigate how DL can be developed and applied to PAM challenges on reef habitats. In Chapter 2, I test the utility of a variety of ML approaches for raw soundscape analysis in controlled experiments using three biogeographically independent reef datasets. Unsupervised learning emerged as a promising method for revealing ecological groupings. Feature embeddings extracted by pretrained deep neural networks outperformed handcrafted eco-acoustic indices, approaching the performance of custom-trained networks while reducing computational and expertise overheads. In Chapter 3, I move from studying raw soundscapes to the detection of individual acoustic events. I identify illegal bomb fishing as a priority use case for acoustic event detection, and work with a local partner to develop and deploy a custom classifier for this in a suspected hotspot within the Indo-pacific. The classifier reduced human review time by orders of magnitude and revealed the frequency of bomb fishing, providing a proof of concept for combining low-cost recorders with DL to assess this destructive practice. In the Chapter 4 I explored pretraining strategies to develop an efficient few-shot transfer learner able to reduce the overheads of developing custom classifiers for acoustic events on coral reefs. I assembled ReefSet, a large annotated coral reef audio library. Models pretrained on ReefSet alone were outperformed by those pretrained on larger avian or generic audio libraries. However, mixing all three domains together yielded the optimum strategy, which was then used to produce the pretrained model, SurfPerch. In Chapter 5, I used SurfPerch and an active learning pipeline to measure ecological outcomes from five biogeographically isolated active reef restoration projects. Using this approach, I identified a suite of sonotype classes linked to ecological functions and rapidly developed classifiers for each. I observed signs of recovery for three of the four functions when compared to healthy and degraded baselines but also note high variability across projects. By demonstrating how DL can be developed and applied to extract insights from tropical coral reef PAM data, this thesis lays the groundwork for more efficient, scalable monitoring of these threatened ecosystems.