North Atlantic right whale detection and localisation using deep learning, spectrogram cross-correlation, and nonlinear Bayesian inversion.

We present a processing routine for marine mammal call detection and localisation, using North Atlantic right whale (NARW) upcalls and moans passively recorded by a sonobuoy grid in July 2018 in the southern Gulf of St. Lawrence, Canada. Right whales are critically endangered and at risk from ship strikes and fishing gear entanglement. We demonstrate the effectiveness of passive acoustic monitoring followed by a three-step analysis as a mitigation tool. The first step detects NARW calls using a publicly available deep learning model, trained on open acoustic datasets, adapted and optimised for the sonobuoy dataset using a smaller adaptation set. The second step applies spectrogram correlation for contact association and time difference of arrival estimates, identifying the same call across multiple sonobuoys. The third step estimates call location via nonlinear Bayesian inversion, assuming direct call propagation, yielding source positions with uncertainty estimates. This three-step analysis effectively detects and localises NARW upcalls, especially from within or near the sonobuoy grid. It enhances information compared to visual surveys and provides realistic uncertainty estimates, supporting its role in conservation and management efforts.