Karlee Zammit

Topic

Detecting Ringed Seal Vocalizations using Deep Learning

School of Earth and Ocean Sciences

Date & location

• Wednesday, August 7, 2024
• 1:30 P.M.
• Clearihue Building, Room B017

Examining Committee

Supervisory Committee

• Dr. Stan Dosso, School of Earth and Ocean Sciences, University of Victoria (Co-Supervisor)
• Dr. William Halliday, School of Earth and Ocean Sciences, UVic (Co-Supervisor)
• Dr. Jon Husson, School of Earth and Ocean Sciences, UVic (Member)
• Dr. Sebastien Fabbro, Department of Physics and Astronomy, UVic (Outside Member)

External Examiner

• Dr. Heloise Frouin-Mouy, Southeast Fisheries Science Centre, NOAA Fisheries

Chair of Oral Examination

• Dr. Simon Devereaux, Department of History, UVic

Abstract

Ringed seals (Pusa hispida), a species of Arctic seal, are listed as a species of special concern in Canada due to a projected loss of habitat caused by the effects of climate change. In order to create effective conservation measures to protect this species, it is important to understand their spatial and temporal distributions. The analysis of passive acoustic monitoring (PAM) data can provide spatial, temporal, and behavioural information through the acoustic detection of a species’ vocalizations. Automated detection methods are necessary to analyze large volumes of PAM data within realistic time-scales. Deep learning (DL) based methods have recently outperformed more traditional methods for the automated detection of marine mammal vocalizations. This thesis develops the first practical automated ringed seal vocalization detector using DL methods. Specifically, ResNet, a convolutional neural network architecture which has shown success for other marine mammal species, is used to perform binary classification of spectrograms containing ringed seal vocalizations. The ensemble model achieves an F1 score above 0.90 for manually-verified segmented spectrograms in both development environments and those unseen during development. The detector was also deployed on two continuous datasets containing data from different years and/or locations than those seen during development, and achieves greater than 0.90 recall, but reduced precision at approximately 0.50 for both datasets. Many of the false positive detections had similar frequency-domain signatures to ringed seal vocalizations. The detector will be available as an open-source command-line-interface tool for researchers to apply to their own data.