Using a novel machine learning approach to alleviate the allometric effect in otolith shape-based species discrimination

Species identification by fish otoliths is an effective and appropriate approach. However, the allometric growth of otoliths can cause discrimination confusion, particularly in juvenile otolith classification. In the Southern Ocean, Chionodraco rastrospinosus, Krefftichthys anderssoni, Electrona carlsbergi, and Pleuragramma antarcticum are frequently caught together in krill fishery as bycatch species. Furthermore, the otolith shape of these four species is relatively similar in juvenile fish, making the identification of fish species difficult. In this study, we tried and evaluated many commonly used machine learning techniques to solve this problem. Eventually, by introducing a triplet loss function (function used to reduce intraspecific variation and increase inter-specific variation), the discrimination confusion caused by the allometric growth of otoliths was reduced. The classification results show that the neural network model with the triplet loss function achieves the best classification accuracy of 96%. The proposed method can help improve otolith classification performance, especially under the context of limited sampling effort, which is of great importance for trophic ecology and the study of fish life history. To this end, we recommend that the effect of allometric growth should be excluded before using the otolith shape to discriminate species.