Publicação

Exploring Sperm Whale Vocalization Patterns through Unsupervised Learning

Ver documento

Detalhes bibliográficos
Resumo:The ocean is a vast and complex environment, home to a wide variety of life forms. Among these creatures is the spermwhale—a species knownfor its size and sophisticated acoustic communication. Studying theirvocalizations provides valuable insights into theirbehavior and ecology. However, manually analyzing these vocalizations is time-consuming and impractical at scale. This thesis explores the use of unsupervised learning techniques to automatically segment, characterize, and cluster sperm whale vocalizations, reducing the need for manual labeling. Initial efforts focused on detecting individual clicks to serve as the basis for segmentation, but challenges with noise, overlapping signals, and inconsistent click properties led to a broader approach. The revised methodology focuses on event-based segmentation and sliding window analysis to capture full vocalization events and group similar sounds based on their properties. The dataset used, provided by the University of the Azores, contains both manually labeled data and the corresponding recordings collected using Digital sound recording TAGs (DTAGs), allowing for evaluation and comparison of clustering performance. The study’s importance lies in its potential applications, such as using clustered sounds as labels for supervised learning tasks or gaining insights into sperm whale behavior by incorporating additional information such as animal depth, location, and current activity.
Autores principais:Fernandes, João Miguel Alves Rodrigues
Assunto:Machine Learning Unsupervised Learning Clustering Signal Processing
Ano:2025
País:Portugal
Tipo de documento:dissertação de mestrado
Tipo de acesso:acesso aberto
Instituição associada:Universidade Nova de Lisboa
Idioma:inglês
Origem:Repositório Institucional da UNL
Descrição
Resumo:The ocean is a vast and complex environment, home to a wide variety of life forms. Among these creatures is the spermwhale—a species knownfor its size and sophisticated acoustic communication. Studying theirvocalizations provides valuable insights into theirbehavior and ecology. However, manually analyzing these vocalizations is time-consuming and impractical at scale. This thesis explores the use of unsupervised learning techniques to automatically segment, characterize, and cluster sperm whale vocalizations, reducing the need for manual labeling. Initial efforts focused on detecting individual clicks to serve as the basis for segmentation, but challenges with noise, overlapping signals, and inconsistent click properties led to a broader approach. The revised methodology focuses on event-based segmentation and sliding window analysis to capture full vocalization events and group similar sounds based on their properties. The dataset used, provided by the University of the Azores, contains both manually labeled data and the corresponding recordings collected using Digital sound recording TAGs (DTAGs), allowing for evaluation and comparison of clustering performance. The study’s importance lies in its potential applications, such as using clustered sounds as labels for supervised learning tasks or gaining insights into sperm whale behavior by incorporating additional information such as animal depth, location, and current activity.