Automated extraction of baleen whale calls based on the pseudo-Wigner–Ville distribution.

This article focuses on an automated method for extracting the time–frequency (TF) contours of low-frequency baleen whale calls using a pseudo-Wigner–Ville distribution (PWVD)-based spectrogram. To address challenges posed by the PWVD’s cross terms, the method incorporates an adaptive threshold in preprocessing, a modified Gaussian mixture probability hypothesis density filter with hybrid tracking (GM-PHD-H) to handle model mismatches, and post-processing steps to remove artificial contours and reconstruct split contours. Simulations and experiments on real acoustic recordings of humpback, bowhead, and blue whales demonstrated that the method achieves high TF resolution with an overall recall of 75% and precision of 78.5%, effectively extracting main call components despite noise and interference. The GM-PHD-H filter, combining Kalman and unbiased finite impulse response (UFIR) tracking cores, improves contour extraction accuracy under clutter and model mismatch conditions, though at increased computational cost.