Using deep learning to track time × frequency whistle contours of toothed whales without human-annotated training...

Area of Science:

• Bioacoustics
• Machine Learning
• Marine Mammal Research

Background:

• Odontocetes (toothed whales) produce whistles with distinct contour shapes.
• Automatic extraction of these whistle contours is crucial for applications like species classification and population estimation.
• Deep learning models show promise but require extensive expert annotation.

Purpose of the Study:

• To develop a cost-effective method for extracting odontocete whistle contours.
• To address the limitations of manual annotation in deep learning models.
• To improve the accuracy of whistle contour extraction using automatically generated pseudo-labels.

Main Methods:

• A novel technique utilizing automatically generated pseudo-labels for training deep learning models.
• Development of an improved loss function to compensate for pseudo-label inaccuracies.
• Evaluation of the technique using pseudo-labels from two different algorithms.

Main Results:

• Standard training methods fail with pseudo-labels; the proposed loss function significantly enhances performance.
• Models trained with pseudo-labels achieved high F1-scores (86.31% and 87.2%).
• Performance is competitive with models trained on a large dataset of expert-annotated whistles (87.47% F1-score).

Conclusions:

• The developed technique effectively learns from cost-efficient pseudo-labels for whistle contour extraction.
• This approach offers a viable alternative to labor-intensive expert annotation.
• The method demonstrates strong performance, enabling broader application in bioacoustic analysis.