Similarity Identification of Southern Resident Killer Whale (SRKW) Call Types Under Sparse Sampling Using Siamese Neural Networks

• Main Author: Zhang, Jie
• Other Authors: Faculty of Computer Science, Master of Computer Science, Hai Wang, Gabriel Spadon De Souza, Carlos Hernandez Castillo, Not Applicable
• Format: Thesis
• Language: English
• Published: 2024
• Subjects: Marine Mammal Conservation; Southern Resident Killer Whale; Acoustic Classification; Few-Shot Learning; Convolutional Neural Network; Data Augmentation; Siamese Network; Similarity measurement; Contrastive Learning; Transfer Learning; Meta-Learning; Killer Whale; Killer whale
• Online Access: http://hdl.handle.net/10222/84495

Southern Resident Killer Whales (SRKW) are highly intelligent marine mammals facing extinction in the North Pacific. These whales emit three types of sounds: clicks, whistles, and pulsed calls, with 43 distinct pulsed call types known as their "dialects". However, due to limited and poor-quality data, only nine call types have sufficient annotated recordings for analysis. To address this challenge, this paper proposes a progressive approach to improve SRKW call type identification. Initially, data augmentation techniques were employed to enhance training data volume, leading to a traditional CNN model achieving 97.8% accuracy on 17 SRKW call types. Subsequently, a Siamese Network model was developed to infer the similarity between call types, achieving remarkable performance with an accuracy of 98.5%. This surpasses the performance reported in current literature on audio multi-class classification using deep learning and machine learning methods. Besides, Siamese Network's generalization ability was evaluated on 9 out-of-training 9 SRKW call types, maintaining noteworthy accuracy and recall but with lower precision, which can be improved through manual review and retraining. This study demonstrates that data augmentation and Siamese Networks are effective strategies for overcoming few-shot learning challenges in SRKW call type identification, achieving robust performance even with limited annotated data.