Addressing the data limitations of deep learning algorithms for the detection of marine mammal species in large acoustic datasets

Bioacoustics, the study of animal sounds, is crucial for ecology and conservation. However, such studies are limited by the time required to process the thousands of hours of recordings they generate. While deep learning (DL) methods have emerged as a powerful solution for some acoustic detection tasks, the complexities of bioacoustics data often violate core assumptions of these models. Here, I present tools that translate DL advancements to real-world bioacoustics applications. First, I demonstrate that annotations made at the level of fixed intervals of a few minutes, rather than labelling each individual call, are not a limitation for training DL models. I design a framework that learns to identify sperm whale (Physeter macrocephalus) click trains (series of short, broadband, regularly spaced clicks) directly from 4-minute recordings without individual clicks labelled, achieving performance comparable to expert auditors. Building on this method, I systematically evaluate how well DL models transfer across recording and labelling conditions. Trained models dropped in performance when deployed on new data. However, pretraining on a large, diverse dataset and fine-tuning with a small selection of new data performed comparably to training on a large sample of the target dataset outright. Active learning, which selects the most informative recordings to audit and add to the training dataset, consistently outperformed all other tested transferability approaches. Finally, I introduce an approach for handling evolving call classification systems through a case study on humpback whale (Megaptera novaeangliae) song units, the smallest components of their songs, designing an expert-in-the-loop framework that achieved performance comparable to full expert annotation on unseen years’ recordings with less than 10% of the expert annotation effort. Through these contributions, this thesis provides novel frameworks for overcoming key mismatches between bioacoustic data realities and DL requirements, addressing the annotation and processing bottleneck that limits bioacoustics studies.