Unsupervised signal decomposition for whale call detection in ocean bottom seismometer data: A non-negative matrix factorization approach

Bryde’s whales are baleen whales regularly documented in China’s coastal waters, forming a nearshore resident population of high ecological and conservation value in the Beibu Gulf. However, passive acoustic monitoring is hampered by high deployment costs and maintenance difficulties, which constrains research on the population’s acoustic behavior. Ocean bottom seismometer (OBS), primarily deployed for geophysical surveys, can incidentally serve as a cost-effective platform for cetacean acoustic monitoring. This study utilized waveform data recorded in March 2025 by OBS from the Beibu Gulf buoy-based ocean seismic observation system. From the OBS-recorded waveforms, 198 call sequences were identified. To extract individual calls from these sequences, we employed Non-negative Matrix Factorization (NMF), a technique widely adopted in computational linguistics for decomposing mixed textual patterns, to decompose mel-spectrograms into interpretable signal components. Results show that NMF automatically extracted 208,135 acoustic signals, classified by spectral peak frequency into three bands with median frequencies of 5.7 Hz, 17.7 Hz, and 37.7 Hz, potentially corresponding to seismic phases of whale calls traveling along different propagation paths. These results demonstrate the potential of NMF for cetacean acoustic monitoring. Beyond this application, as a general framework for decomposing mixed signals into spectrally distinct components, NMF lends itself naturally to seismological source-separation tasks, such as separating overlapping seismic phases and detecting repeating earthquakes.