Otolith identification using a deep hierarchical classification model

The diet of seabirds can yield important insights into the status of economically and ecologically important fish. By analyzing the otoliths found in the birds' droppings, researchers can observe which fish the birds eat in which abundances. However, identifying the species based on an otolith image is quite labor-intensive and requires particular expertise. In this work, we show that a deep convolutional neural network can identify six fish species with high accuracy. We show that this deep learning approach outperforms more traditional methods and is also more accessible to set up in practice. By exploiting the hierarchy in the species labels, we impose a structure on the prediction probabilities, leading to a remarkable improvement compared to a conventional artificial neural network. Importantly, we can attain good results using only a modest dataset, demonstrating that such approaches are feasible for small-scale and specialized projects.

Automated summary

Studying the diet of seabirds can provide valuable insights into the status of important fish species, and a deep convolutional neural network has shown high accuracy in identifying fish species based on otolith images, outperforming traditional methods and being more practical to implement.