Advancing passive acoustic ship detection in the Western Canadian Arctic: Signal processing and deep learning approaches

The retreat of Arctic sea ice is driving an increase in vessel traffic and associated underwater noise, which interferes with the frequency bands used by Arctic marine mammals. Detecting co-occurring vessel noise and marine mammal vocalizations in passive acoustic monitoring (PAM) data can help to assess their adverse impacts and guide mitigation strategies. This paper proposes two ship noise detection techniques: a modified variant of the Frequency Amplitude Variation (FAV) method, MFAV, which integrates signal processing with a simple statistical threshold to enhance both interpretability and detection performance; and a convolutional neural network (CNN) model specifically trained to advance ship detection in the Canadian Arctic. Comparative analysis of our PAM test dataset from the western Canadian Arctic, based on peak F1-scores, demonstrates that the CNN model generalizes well to unseen sites and, with one exception, consistently outperforms both MFAV and FAV by 1%-8%, maintaining scores above 91%. Furthermore, MFAV improves the detection of boats by up to 22% and of larger ships by 6%. The developed methods are publicly available as an open-source tool on GitHub, contributing to the advancement of acoustic vessel monitoring techniques in Canadian Arctic waters in support of conservation efforts aimed at protecting Arctic marine mammal habitats.