Evaluation of a signal processing and machine learning framework to detect and classify shell-crushing predation events

Contributions of large mobile predators to shellfish mortality are poorly understood, limiting our scientific understanding of how they control prey populations in aquatic ecosystems. In this study, we present an advanced methodology for detecting and classifying shell-crushing events from acoustic recordings of the whitespotted eagle ray ( Aetobatus narinari ) consuming various hard-shelled mollusks. Data were acquired primarily from controlled captive experiments, where we recorded 4 individuals feeding on a representative bivalve and two gastropod species. Using a matched filter coupled with time-frequency handcrafted features, such as gammatone spectral coefficients (GTCC) and deep convolutional neural networks (CNN) as unsupervised feature extractors, we developed a comprehensive method for detecting and classifying shell-crushing sounds. Additionally, we validated our detection and classification scheme based on a field experiment in natural conditions where we both simulated shell fracture of hard clams, and recorded whitespotted eagle rays naturally consuming hard clams and other shelled organisms. The integration of machine learning techniques, including conventional algorithms such as Random Forest, Support Vector Machines, and deep learning Long Short-Term Memory (LSTM) networks, enabled the characterization of these bioacoustic events in terms of prey species being consumed. The findings demonstrate superior ability of the LSTM classifier and the similar performance of GTCC and CNN features in classifying shell-crushing events. Additionally, this study underscores the possibility of detecting and classifying shell-crushing events in the wild, improving our understanding and monitoring of marine ecological interactions. Such developments improve our capacity to study the foraging ecology of large mobile predators, their natural effects on marine benthic communities as well as interactions with shellfish restoration programs. • Sounds produced by the fracture of mollusk shells during predation events are unique and detectable. • A machine learning algorithm was built to classify the fracture sounds in terms of mollusk species. • Although most training sounds were collected in a tank, the algorithm shows potential to identify feeding events in the wild.