Cocoa Beans Roasting Classification using Hybrid Multi-Objective Golden Eagle Optimiser and Graph Convolutional Networks

Abstract

The roasting process greatly influences cocoa beans’ quality and antioxidant potential, transforming phenoliccompounds and flavour precursors. However, the conventional method of evaluating roasting levels relies on subjectivehuman perception, which can lead to inconsistencies in quality control. This study proposes an intelligent and objectiveclassification system for cocoa bean roasting levels (light, medium and dark) based on RGB images, using a Hybrid Multi-Objective Golden Eagle Optimiser approach integrated with Graph Convolutional Networks (HMOGEO–GCN). The studyuses 1000 images for each roasting class. The ResNet18 model is a feature extractor optimised through multi-objective GEOto balance accuracy, computational cost and convergence speed. GCN strengthens population topology learning for a moreeffective global search. Using a 10-fold cross-validation, experimental results show that HMOGEO–GCN achieves superiorperformance, with an average accuracy of 90.5% ± 0.8, an F1-score of 0.913 and an AUC of 0.904. This surpasses theperformance of standard CNNs and single optimisation approaches. Furthermore, the classification results demonstrate astrong correlation with chemically reported antioxidant activity and total polyphenol content, showing the potential of thismodel for predictive quality control in the cocoa roasting process. This approach contributes to developing deep learning forfood analysis and supports implementing intelligent, precision cocoa roasting systems in industry.