Researchers have trained a machine learning model to identify the diet of marine predators by analyzing the acoustic signatures of them crushing shells. The tool provides a non-invasive method to track predator-prey dynamics, a critical metric for marine conservation that previously required intrusive or visually obstructed observation techniques.
Overcoming underwater observation limits
Tracking what marine animals eat underwater presents distinct physical challenges. Cameras mounted on predators often capture little more than clouds of displaced sand, while extracting stomach contents requires capturing and flushing the animal, which degrades the evidence. Matt Ajemian, an assistant research professor at the Harbor Branch Oceanographic Institute at Florida Atlantic University, said the team shifted focus to the sounds animals produce when breaking open shells.
Training the model on eagle rays
The team tested the algorithm on whitespotted eagle rays in both controlled tanks and the ocean off the Florida coast. They recorded the animals feeding on known quantities of clams and snails while wearing biologgers that captured audio and video. The data revealed distinct processing times for different prey.
Rays take longer to eat clams because they must sift out the shells, whereas snails detach quickly once the single attachment point breaks. The model learned to filter out background ocean noise and classify the specific prey based on these crushing sounds. "A lot of animals out there, particularly marine animals, have the unique ability to crush shells open," Ajemian said. "But we don't know how much they eat and what they feed on. So we wanted to see if we could remotely detect an animal feeding on a clam versus a gastropod."
Computing requirements and future applications
The researchers found that highly complex neural networks were unnecessary for accurate detection. Simpler algorithms required a fraction of the computing resources while delivering nearly identical accuracy. "The method that required the most computing power wasn't necessarily the method or approach that yielded the best results," Ajemian said. "There were ones that would take a fraction of the computing power that were pretty darn close."
This efficiency means the tool can run on standard hardware, making it accessible for broader AI for Science & Research applications. The team plans to extend the model to include crabs and apply the methodology to long-term acoustic recordings from stationary ocean stations. This approach will help scientists map exactly when and where shellfish are consumed across different habitats, supporting more targeted research in marine ecology.
Why this matters for science and research professionals
Acoustic monitoring offers a scalable alternative to visual tracking in environments where cameras fail. For scientists designing ecological studies, integrating passive audio sensors with lightweight machine learning models can yield high-resolution dietary data without disturbing the subjects. Evaluating simpler algorithmic architectures early in the design process can also reduce computational costs while maintaining analytical rigor.
Your membership also unlocks:
