Detecting, classifying, and counting blue whale calls with Siamese neural networks

Abstract

The goal of this project is to use acoustic signatures to detect, classify, and count the calls of four acoustic populations of blue whales so that, ultimately, the conservation status of each population can be better assessed. We used manual annotations from 350 h of audio recordings from the underwater hydrophones in the Indian Ocean to build a deep learning model to detect, classify, and count the calls from four acoustic song types. The method we used was Siamese neural networks (SNN), a class of neural network architectures that are used to find the similarity of the inputs by comparing their feature vectors, finding that they outperformed the more widely used convolutional neural networks (CNN). Specifically, the SNN outperform a CNN with 2% accuracy improvement in population classification and 1.7%-6.4% accuracy improvement in call count estimation for each blue whale population. In addition, even though we treat the call count estimation problem as a classification task and encode the number of calls in each spectrogram as a categorical variable, SNN surprisingly learned the ordinal relationship among them. SNN are robust and are shown here to be an effective way to automatically mine large acoustic datasets for blue whale calls. V

Passive acoustic monitoring of killer whales (Orcinus orca) reveals year-round distribution and residency patterns in the Gulf of Alaska

“…Recordings were processed using the whistle and moan detector in the passive acoustic monitoring software PAMGuard (version 1.15.17) to automatically identify likely cetacean vocalizations 91 . The PAMGuard detector was tuned to prioritize sensitivity over accuracy or specificity (i.e., increased false positive rate, decreased false negative rate 92 to minimize the probability of missing killer whale encounters. Recordings with PAMGuard detections were then aurally and visually inspected in Audacity® (version 2.3.2 93 ), fast Fourier transform size 1024 with a Hann window) to confirm presence or absence of killer whale pulsed calls and/or whistles.…”

Bioacoustic Event Detection with Self-Supervised Contrastive Learning

“…Following the denoising layer is the convolutional model, consisting of Conv1d layers with batch normalization (Batch-Norm1d) and LeakyReLU activation. We use a 4-layer network with kernel sizes [8,6,4,4] and strides [4,3,2,2], corresponding to a hop length of 48 samples (i.e. 1ms) and a receptive field of 136 samples (i.e.…”

“…Large-scale studies often employ Convolutional Neural Network (CNN) architectures to carry out supervised detection or classification tasks across diverse species. These include detection of humpback whale vocalizations in passive acoustic monitoring datasets 5 ; detection, classification, and censusing of blue whale sounds 6 ; avian species monitoring based on a CNN trained to predict the species class label given input audio 7 ; presence indication of Hainanese gibbons using a ResNet-based CNN 8 ; and, recently, detection and classification of marine sound sources using an image-based approach to spectrogram classification 9 . However, such supervised learning approaches remain limited in their scope, hindering their capacity to be deployed in real-time data processing pipelines.…”

Deep learning algorithm outperforms experienced human observer at detection of blue whale D‐calls: a double‐observer analysis

“…Thus, having an automated detector for blue whale D-calls would facilitate a more complete knowledge of blue whale spatio-temporal occurrence and acoustic behaviour from the hundreds of thousands of hours of underwater recording that have already been analysed for blue whale song (e.g. Branch et al, 2021;Sirovi c et al, 2017Sirovi c et al, , 2018. Automated detectors for D-calls could also prove useful during voyages that employ real-time tracking to locate these endangered animals for further study (e.g.…”

“…In recent years, machine learning algorithms, such as deep‐learning, have shown great promise for fast and accurate analysis of PAM data (Bergler et al., 2019; Madhusudhana et al., 2021; Rasmussen & Širović, 2021; Shiu et al., 2020; Zhong et al., 2021). In other domains, these methods have developed to the point that they exceed the performance of human observers (e.g.…”