Using context to train time-domain echolocation click detectors.

Area of Science:

• Marine Biology
• Acoustics
• Machine Learning

Background:

• Automated detection of toothed whale echolocation clicks is crucial for ecological monitoring.
• Previous methods often require extensive manual validation, limiting scalability.

Purpose of the Study:

• To demonstrate the efficacy of human-in-the-loop processes for creating large training datasets for machine learning.
• To develop and evaluate a machine learning system for automated detection and classification of toothed whale echolocation clicks.

Main Methods:

• A permissive energy-based detector identified potential echolocation clicks.
• A machine-assisted quality control process, leveraging contextual cues, curated a dataset of over 57,000 clicks.
• Feed forward neural networks were trained on subsets of this data.
• The trained networks detected over 850,000 echolocation clicks, validated using the same quality control process.

Main Results:

• The neural network architecture demonstrated robust performance across various contexts.
• Evaluation against a separate dataset, collected five years apart and over 600 km distant, confirmed reliability.
• The system successfully identified echolocation bouts missed by human analysts.
• Classifier errors primarily involved anthropogenic noise, not included as negative training examples.

Conclusions:

• Human-in-the-loop approaches are effective for constructing large, high-quality machine learning training sets.
• The developed system offers a scalable and accurate method for toothed whale echolocation click detection.
• The system shows high performance in real-world scenarios, with low false positive rates in the absence of anthropogenic noise.