DeepWhaleNet: A Climate Change-Aware FFT-Based Neural Network for Underwater Passive Acoustic Monitoring

• Main Author: Ryan Rasmussen, Nicholas
• Format: Text
• Language: unknown
• Published: USD RED 2024
• Subjects: Artificial Intelligence; Bioacoustics; Climate Adaption; Conservation; Deep Learning; Signal Processing; Acoustics; Dynamics; and Controls; Artificial Intelligence and Robotics; Biology; Blue whale
• Online Access: https://red.library.usd.edu/diss-thesis/264; https://red.library.usd.edu/context/diss-thesis/article/1256/viewcontent/Nicholas_Rasmussen_digitally_accessible_1.pdf; https://red.library.usd.edu/context/diss-thesis/article/1256/filename/0/type/additional/viewcontent/abstractPage.pdf; https://red.library.usd.edu/context/diss-thesis/article/1256/filename/1/type/additional/viewcontent/3_Committe_Signature_page.docx.pdf

In the face of escalating climate threats, the conservation of whale species has become increasingly critical. Traditional acoustic monitoring methods, burdened by extensive pre-processing and post-processing, need more adaptability and efficiency for effective marine mammal surveillance. This study introduces DeepWhaleNet, a novel deep-learning framework tailored for Underwater Passive Acoustic Monitoring (UPAM). DeepWhaleNet is designed to streamline whale detection by directly analyzing raw log-power spectrograms, thus extracting essential acoustic features to conserve these endangered species. The framework employs an extensive short-time Fourier transform (STFT) for input processing and a customized ResNet-18 architecture for classification, distinguishing whale vocalizations from ambient noise and accurately identifying their time-frequency signatures. Evaluation of DeepWhaleNet reveals its superiority over conventional models, with an 8.3% increase in the F-1 score and a 21% improvement in average precision for binary relevance. Furthermore, the model’s versatility and precision in species-specific sound detection are confirmed through an ablation study, achieving a 99.1% recall rate for Blue Whale calls. This research enhances the benchmark performance for whale call detection. It paves the way for future integration of advanced machine learning techniques, such as active learning, to further refine the reliability of dataset annotations and the spatial specificity of whale vocalizations within spectrograms.