The application of neural networks to classify dolphin echolocation clicks

Abstract

Passive acoustic monitoring (PAM) is a common approach to monitor marine mammal populations, for species of dolphins, porpoises and whales that use sound for navigation, feeding and communication. PAM produces large datasets which benefit from the application of machine learning algorithms to automatically detect and classify the vocalisations of these animals. We present a deep learning approach for the classification of dolphins’ echolocation clicks into two species groups in an environment with high background noise. We compare the use of Convolutional Neural Networks (CNN) and Recurrent Neural Network (RNN), in which we feed the models the raw waveform data and spectrograms. We show that both models perform well, with the highest performance achieved by a CNN fed with spectrograms (F1 score 97 %) and an RNN fed with raw data (F1 score 96%) fitted with Gated Recurrent Units (GRU). We recommend the use of such models to classify echolocation clicks in marine environments where background noise levels exhibit high spatial and temporal variance. In particular, the RNN showed excellent performance, while being fed with raw data, in terms of reduced processing time and storage. Deep learning automatically extracts effective features from the raw waveform in the training process through multiple layers of the model, without the need to rely on feature extraction in a separate pre-processing step.

Bionic study of distance-azimuth discrimination of multi-scattered point objects in bat bio-sonar

“…There are many organisms in nature that rely on echolocation, such as bats, dolphins, whales, etc all of which employ sound to 'see' to realize detection, positioning, communication, predation and other functions [1][2][3][4]. Specifically, bats emit highfrequency harmonic signals by twisting their mouths or noses.…”

A real-time passive acoustic monitoring system to detect Yangtze finless porpoise clicks in Ganjiang River, China

“…Originally, researchers marked the target signals by eye, manually checking the spectrogram of the acoustic data (Goold and Jefferson, 2002), which is a large amount of work and greatly dependent on the operator's subjective judgment. Thus, some automatic programs have been developed, such as automatic filters with several parameters (such as the sound pressure level of clicks after band-pass filtering or the inter-click time intervals) based on event data loggers (Thomsen et al, 2005;Kyhn et al, 2008;Sven et al, 2008;Kimura et al, 2010.;Bailey et al, 2010) or spectrogram classification algorithms based on a convolutional neural network model to separate clicks from other noise (Mellinger, 2021;Seydi et al, 2022). However, the automatic filters were mainly built on event data loggers and important acoustic parameters, such as the center frequency of clicks or 3-dB bandwidth, are not included in the filters (Kimura et al, 2010;Kyhn et al, 2013;Fang et al, 2015).…”