Interpretability and identification of dimorphism in morphological indexes of Larimichthys crocea based on machine learning models

Liguo Ou, Linlin Lu, Weiguo Qian, Bilin Liu,
Interpretability and identification of dimorphism in morphological indexes of Larimichthys crocea based on machine learning models,
Fisheries Research,
Volume 288,
2025,
107475,
ISSN 0165-7836,
https://doi.org/10.1016/j.fishres.2025.107475.
(https://www.sciencedirect.com/science/article/pii/S0165783625002127)
Abstract: The morphological indexes serve as a critical biological foundation for analyzing species dimorphism, play a pivotal role in population dynamics models and species assessments, and provide valuable, accurate, and cost-efficient biological information. Dimorphism identification holds significant importance for the conservation and sustainable development of Larimichthys crocea resources. Therefore, this study aims to validate the dimorphism effects of various morphological indexes using interpretable machine learning techniques and evaluate model performance and deviation in automatic identification. First, data visualization, significance analysis, correlation analysis, and principal component analysis (PCA) were applied to otolith morphology (OM) indexes and fish body morphology (FM) indexes. Then, the SHAP (SHapley Additive exPlanations) method of machine learning was used to analyze the importance of different morphological indexes and output the morphological indexes of importance. Finally, different machine learning models were used to analyze the identification performance and deviation of Larimichthys crocea dimorphism. The experimental results demonstrate that the SHAP method effectively prioritizes the importance of different morphological indexes, with the importance of OM indexes primarily concentrated in the sulcus. Within the machine learning models, OM indexes achieved a peak identification rate of 71 % (Random Forest), whereas FM indexes reached a maximum identification rate of 65 % (Random Forest and Support Vector Machine). The comparative analysis of the average effects of different models, including evaluation metrics and learning curves, demonstrates that OM indexes outperform FM indexes in terms of identification performance. The application of machine learning models not only enables a comprehensive analysis of the dimorphism in Larimichthys crocea but also offers effective strategies for the conservation of Larimichthys crocea resources and their associated biodiversity.
Keywords: Machine learning; SHAP; Larimichthys crocea; Dimorphism; Otolith; Morphological indexes