Tagging ocean soundscapes with machine learning and topological approaches

Abstract

With a multitude of sources simultaneously contributing to the ambient ocean soundscape, it is important to disaggregate noise source contributions for better monitoring and assessing noise impacts. Our work adopts a modular set of neural networks and topological pipelines to analyze multiple representations of time-series samples of the ambient ocean noise. Specifically, a convolutional neural network (CNN) is used to broadly tag the presence of local shipping and marine mammal vocalizations from spectrogram representations. Multiple approaches are investigated using topological data analysis to predict a class label of the ship hull type or specific marine mammal species using the audio data as input. The datasets used in training, testing, and validation comprise hydrophone recordings across varying sensors and ocean environments to address the wide variation in ambient background noise and propagation paths. Algorithm performance is characterized by classification accuracy on labeled data. This approach demonstrates a generalized tagging cabilitility of the presence of marine mammals and local shipping activity.