Automatic characterization of beluga whale populations by combining passive acoustics and machine learning

Irène Roca, Tristan Cotillard, Xavier Secheresse, Jaclyn Aubin, Marie-Ana Mikus, Valeria Vergara, Sébastien Gambs, Angéliuqe Dupuch, Véronique Lesage, Robert Michaud, Cristiane C. A. Martins, Samuel Turgeon, Clément Chion

The beluga (Delphinapterus leucas) population inhabiting the waters of the St. Lawrence Estuary (SLEB) was in decline and counting ~900 individuals when last assessed in 2012. Multiple factors have been identified as responsible, including marine traffic and the related noise. Being a highly social and vocal species, passive acoustic monitoring has the potential to deliver, in a non-invasive and continuous way, real-time information on SLEB spatio-temporal habitat use, which is crucial for their conservation. We develop machine learning pipelines to automatize the analysis of passive acoustic data and provide standard and accurate estimations of animal presence, abundance and vocal repertoire. We focus on a high residency area within SLEB summer habitat (Baie Sainte-Marguerite –BSM) where recent research has shown that beluga presence-time decreases with ship transit. Using spectral power density of well-chosen wavelets as input features, we train a binary classifier to first detect periods with SLEB acoustic presence. Then, a double thresholding technique is applied to extract regions-of-interest (ROIs) within the presence spectrograms and feed the resulting ROI images to a convolutional neural network classifier to discriminate vocalization types. To separate SLEB’s abundance classes we feed relevant input features aggregated over large temporal windows to a deep learning classification algorithm. Preliminary results showed that models achieved an overall accuracy of 95%, 85% and 70% to discriminate SLEB’s presence, call type and abundance, respectively. Accurate continuous estimations of SLEB abundance and vocal activity at BSM could provide valuable information to quantify habitat use and disentangle the functional relation between the observed decrease in SLEB presence-time and ship transit. Understanding this relationship is crucial for the co-construction of measures to mitigate the impacts of vessel traffic that are effective for the recovery of the SLEB and realistic for the industry