Publicação

Automatic detection of beaked whale echolocation clicks via convolutional neural networks

Ver documento

Detalhes bibliográficos
Resumo:In ecological studies, over recent decades, biological datasets have increased rapidly, both in size and complexity. This emphasises the necessity for ecologists to have practical tools to analyse this abundance of data and an understanding of modern techniques. Here, machine learning plays an important role since it automates and streamlines the process, enhancing efficiency in analysing and understanding the large amount of data available. Using data collected primarily from digital acoustic tags (DTAGs) attached to Blainville’s beaked whales (Mesoplodon densirostris), supplemented by additional information from Cuvier’s beaked whales (Ziphius cavirostris), the goal of this thesis was to develop a modern machine learning model capable of automatically identifying the distinct echolocation clicks emitted by these species using deep learning techniques via convolutional neural networks (CNNs). Two distinct experiments were conducted. Firstly, to evaluate the models’ classification capabilities. Secondly, an existing echolocation click detector tool was compared to the CNN models developed in this study. Each experiment included multiple scenarios, consisting of different dataset configurations and objectives, that were assessed using accuracy, recall, precision, and F1-score. The CNN developed in this study, designed as a binary classifier for detecting the presence or absence of Cuvier’s beaked whales, achieved an F1-score of 93.28% when applied to 15.8 hours of data, which corresponded to 27568 correctly classified audio segments and 802 incorrectly classified. While the CNN model developed for Blainville’s beaked whales achieved an F1-score of 84.12% when applied to 3.4 hours of data, which corresponded to 5544 correctly classified audio segments and 521 incorrectly classified. The difference in performance between Cuvier’s and Blainville’s beaked whale models could be attributed to data scarcity. By developing these CNN models, the aim was to identify echolocation clicks and, furthermore, to provide a steppingstone for accurately estimating population densities, which can be done by resorting to methodologies such as cue counting.
Autores principais:Gueifão, Tomás Lourenço Marques Vicente
Assunto:Bioacústica Redes Neurais Convolucionais Machine Learning Deep Learning Baleia de bico Trabalhos de projeto de mestrado - 2024
Ano:2024
País:Portugal
Tipo de documento:dissertação de mestrado
Tipo de acesso:acesso aberto
Instituição associada:Universidade de Lisboa
Idioma:inglês
Origem:Repositório da Universidade de Lisboa

Registos relacionados

Registos relacionados