An acoustic monitoring method for marine mammals based on the integration of dual-comb multi-heterodyne interferometry and lightweight deep learning models

Abstract To overcome the limitations of traditional acoustic methods in sensitivity, classification accuracy, and deploy ability, this study proposes a novel framework that combines dual-comb multi-heterodyne interferometry (DCMHI) with deep learning. The system enables high-speed, high-sensitivity detection of sound pressure variations across frequencies using optical frequency combs (OFCs). A Mel-spectrogram dataset based on real marine mammal vocalizations was constructed, with multiple audio augmentation techniques employed to enhance data diversity and model generalization. Meanwhile, a Lightweight Marine Mammal Classification Network (LMMC-Net) model was designed, which integrates the Spatial-Channel Semantic-Guided Block (SCSGB) and the Linear Manhattan Self-Attention Block (LMASAB) to enable efficient acoustic feature extraction and accurate classification. Experiments show that the system achieves a minimum detectable sound pressure level (MDSPL) of 97.68 dB re 1 μPa for a 6 kHz pulse when the output signal-to-noise ratio (SNR) reaches 15 dB. Data augmentation improves classification accuracy by 2.6%. On the Mammalian Mel-Spectrogram Dataset (MMS-31 dataset), LMMC-Net outperforms YOLOv12 by 1.1% in Top-1 accuracy and 21.63% in inference speed, while reducing parameters by 41.62%, demonstrating strong potential for edge deployment.