Techniques to Overcome Data Scarcity in Deep Learning for Passive Acoustic Monitoring of Marine Mammals

Abstract

Passive Acoustic Monitoring (PAM) is a useful technique for monitoring marine mammals. However, the large volume of data collected through PAM systems make automated algorithms for detecting and classifying sounds essential. Deep learning algorithms have shown great promise in recent years, but their performance is limited by insufficient amounts of annotated data for training the algorithms. Our work examines several machine learning techniques to overcome data scarcity in a single and multi-domain scenarios, where each domain is a different underwater acoustic environment. We first investigate the benefits of augmenting training datasets in a single domain with synthetically generated samples when training a deep neural network for the classification of marine mammals. We apply two acoustic data augmentation techniques, SpecAugment and Mixup, on PAM data to improve the network`s performance. Next, we address the challenge of data scarcity in a multi-domain context through transfer learning, a machine learning concept whereby knowledge from a source domain is transferred to a target domain. Specifically, we considered two different underwater acoustic environments as the source and target domain. We develop a more robust deep neural network model for the classification of marine mammals by incorporating knowledge from two different domains. Lastly, we confront data scarcity in a scenario where no annotated data is available for training deep learning models. In this context, we explore the artificial generation of synthetic marine mammal vocalizations, integrating real acoustic properties from the underwater environment to create datasets for training deep neural networks in detecting and classifying real marine mammal vocalizations. We evaluate the performance of all three approaches and compare the results with baseline models. We demonstrate that the proposed approaches provide useful and effective solutions in scenarios of data scarcity under diverse and variable conditions.