Automated Detection and Classification of Marine Species Vocalizations Using a YOLO-Based Deep Learning Framework

Area of Science:

• Marine Biology
• Acoustics
• Artificial Intelligence

Background:

• Underwater acoustic environments are complex, with overlapping natural and anthropogenic signals.
• Effective marine species detection and classification are crucial for ecological monitoring and understanding human impacts.
• Limited availability of continuous, well-annotated multi-species datasets hinders deep learning model development.

Purpose of the Study:

• To propose a deep learning framework for automatic detection and classification of marine species vocalizations.
• To address the challenge of limited annotated datasets by creating synthetic monitoring data.
• To evaluate the model's performance in realistic, overlapping acoustic scenarios.

Main Methods:

• A deep learning framework inspired by the YOLO (You Only Look Once) architecture was developed.
• Synthetic monitoring datasets were generated by combining single-species vocalizations.
• Augmentation techniques, including CutMix, were applied to enhance dataset diversity and robustness.
• The model was tested under both non-overlapping and overlapping signal conditions.

Main Results:

• The proposed model achieved strong performance in detecting and classifying marine species vocalizations.
• The framework demonstrated stable performance even in complex, overlapping acoustic scenarios.
• The study validates the effectiveness of YOLO-inspired architectures in diverse underwater acoustic conditions.

Conclusions:

• Deep learning models, particularly YOLO-inspired architectures, can effectively monitor marine species in challenging acoustic environments.
• Synthetic data generation and augmentation are viable strategies to overcome dataset limitations.
• Further research incorporating long-term field recordings is recommended to enhance reliability.