Abstract
This study proposes a Variational AutoEncoder (VAE)-based Quality Controlled (QC) generative augmentation framework for North Atlantic right whale (NARW) upcall detection. Existing generative augmentation methods can generate synthetic samples; however, they do not provide a sample-level criterion for determining whether each generated sample is a positive sample that contributes to improved detector performance or a synthetic outlier that should be removed. This study learns manually extracted upcall frequency contours using a 1D-VAE and evaluates generated contour candidates in a 10-dimensional acoustic morphology feature space. The QC score is computed with respect to the reference distribution of manually extracted real upcall contours, and stochastic acceptance probabilities for borderline samples around the hard threshold are calibrated using same-call manual re-extraction variability. QC-passed contours are converted into detector training spectrograms using smooth amplitude modulation and Gaussian noise injection based on SNR statistics. Using 30,000 acoustic segments from the Kaggle NARW dataset, Original, Denoising Diffusion Probabilistic Models (DDPM), Contour-VAE without QC, and VAE-QC conditions were compared under the same detector and augmentation budget. VAE-QC with α = 0.97 achieved the highest mean 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 further showed that VAE-QC suppresses morphology outlier tails observed in unfiltered generation. These results indicate that the key factor in generative augmentation is the QC process that defines feature boundaries useful for detector learning and selects synthetic positive samples accordingly.
Showing the abstract — retrieve the full paper via the Exa API.