Acoustic analysis of big ocean data to monitor fish sounds

•The challenging task of fish sound monitoring from single channel audio has been addressed.•The proposed feature sets are constructed based on both local and contextual information.•The proposed feature selection is taken into account to reduce non-distinctive features.

This paper presents a novel framework for monitoring fish sounds based on acoustic analysis of noisy big ocean data. The proposed method involves multiresolution acoustic features (MRAF) extraction and RPCA (robust principal component analysis) based feature selection for monitoring of natural fish sounds produced in situ by the plainfin midshipman (Porichthys notatus); here, we investigate this fish's grunts, growls and groans. Both local and contextual information are exploited by MRAF, while sparse components of the MRAF matrix obtained through RPCA is found to be more robust to overlapping low-frequency spectral contents among different classes. The simulation results obtained from real-recorded ocean data reveal the advantages of the proposed scheme for monitoring underwater soundscapes and determining a variety of fish sounds in natural marine habitats.