Acoustic detection of marine mammals by deep learning

• Main Author: Casals I Salvador, Marc
• Other Authors: Universitat Politècnica de Catalunya. Departament de Teoria del Senyal i Comunicacions, Hernando Pericás, Francisco Javier
• Format: Master Thesis
• Language: English
• Published: Universitat Politècnica de Catalunya 2023
• Subjects: Àrees temàtiques de la UPC::Enginyeria de la telecomunicació::Processament del senyal; Deep learning (Machine learning); Whales > Classification; Deep Learning; Acoustic Classification; Accuracy; Spectral Subtraction; Data Augmentation; CNN; mel-spectrogram; Residual Networks; Aprenentatge profund; Balenes > Classificació; Blue whale
• Online Access: http://hdl.handle.net/2117/403156

In oceanography, specifically in the study of whales, the classification of the different species is an essential task for passive acoustic observation. It is fundamental for the tracking of some species that have become endangered over time due to the human hand. This Master Thesis focuses on the categorization of two species: the Blue Whale and the Finn whale. These are closely related and both of them have multiple vocalization types, making their categorization challenging. This work makes usage of Deep Learning models, specifically Deep Convolutional Neural Networks. The main architectures that have been used are LeNet5 and ResNet50. These architectures are encompassed with some Data Augmentation and Signal-Processing techniques that make an improvement of their performance, such as time and frequency masking or spectral subtraction. The results showed that ResNet50 was the most capable architecture, and combined with spectral subtraction and data augmentation, it was possible to make an upstanding improvement of the baseline.