Quality-Controlled Generative Augmentation for North Atlantic Right Whale Upcall Detection Using Contour 1D-VAE

Area of Science:

• Bioacoustics
• Machine Learning
• Marine Mammal Research

Background:

• Generative augmentation methods for acoustic detection can produce synthetic outliers.
• Existing methods lack sample-level criteria to distinguish useful synthetic data from outliers.
• North Atlantic right whale (NARW) upcall detection requires high accuracy to aid conservation efforts.

Purpose of the Study:

• To propose and evaluate a Quality Controlled (QC) generative augmentation framework based on Variational AutoEncoder (VAE) for NARW upcall detection.
• To develop a sample-level criterion for assessing the quality of generated acoustic samples.
• To improve the performance of NARW upcall detectors by effectively utilizing synthetic data.

Main Methods:

• A 1D-VAE was used to learn manually extracted upcall frequency contours.
• A QC score evaluated generated contour candidates in a 10-dimensional feature space against real upcall distributions.
• QC-passed contours were converted into training spectrograms with amplitude modulation and Gaussian noise.

Main Results:

• The VAE-QC framework achieved the highest mean Area Under the Curve (AUC) of 0.901, outperforming original training (0.802), DDPM (0.845), and Contour-VAE without QC (0.810).
• Feature distribution and QC-score analyses demonstrated that VAE-QC suppresses morphology outlier tails.
• The study confirmed the critical role of the QC process in generative augmentation for detector learning.

Conclusions:

• The proposed VAE-QC framework significantly enhances NARW upcall detection performance.
• A robust QC process is essential for generative augmentation, enabling the selection of high-quality synthetic positive samples.
• This approach offers a promising direction for improving acoustic monitoring of endangered marine species.