Applications of Deep Convolutional Neural Networks to Passive Acoustic Monitoring of Baleen Whales

Abstract

Research into automated detection and classification systems (DCS) of marine mammal vocalizations in acoustic recordings is expanding internationally due to the necessity to analyze large collections of data for conservation purposes. This work discusses historical implementations of marine mammal DCS and introduces more recent developments using deep learning and convolutional neural networks (CNNs). A novel application of region-based CNNs (R-CNNs) is used for the development of a DCS that is capable of detecting the vocalizations of endangered baleen whales in both time and frequency. A state-of-the-art approach to semi-supervised learning is adapted for the task of spectrogram classiőcation in order to address the issue of data scarcity commonly found in passive acoustic monitoring (PAM) and increase the performance of the aforementioned systems by as much as ten percent. Finally, an operational marine mammal DCS is developed to address real-world constraints like data set variance, computational limitations, and over the-air model updates. This work contributes to advancing marine mammal vocalization detection and supports conservation efforts through PAM.