Age recognition of a semi-pelagic fish (Carangiformes: Carangidae) using a Swin Transformer and Gaussian Process Classifier with otolith images

Abstract

Accurate age determination of commercially important fish species is essential for sustainable fisheries management and stock assessment. However, traditional methods relying on the manual counting of otolith annuli are labour-intensive, time-consuming, and subject to significant inter-reader variability. This study introduces SwinGPC-AgeRecognitioNet, a hybrid deep learning framework designed for efficient automated age estimation in the Mediterranean horse mackerel (Carangidae: Trachurus mediterraneus), to address these challenges. The proposed architecture synergises Swin Transformer-based feature extraction with a Gaussian Process Classifier (GPC) to capture global structural patterns while providing robust probabilistic predictions. The methodological workflow integrates three critical stages: (1) high-level feature extraction via Swin Transformer; (2) discriminative feature selection using Recursive Feature Elimination; and (3) hyperparameter-optimised classification via GPC. Experimental evaluations on a dataset of 1231 otolith images reveal that the proposed model consistently outperforms Convolutional Neural Networks architectures (e.g., VGG, ResNet), achieving accuracies of 88.66 % in multi-class classification and up to 94.33 % in binary tasks. These findings highlight the potential of SwinGPC-AgeRecognitioNet as a scalable, high-precision tool for fisheries science, offering a reliable alternative for data-driven resource management.