When dealing with a multi-class classification, we must take into account both the overall performance and the performance of each individual class. While some classes may have straightforward patterns, others may have more complex ones. Our goal is to improve the predictive performance by applying the stacking technique to the cocoa bean cut-test image classification. We take advantage of a diverse set of model in a stack to reduce prediction errors for some classes that are difficult to classify in order to improve performance. The overall accuracy, precision, recall, and f1-score are used as evaluation
metrics, and the results demonstrate that the stacked model achieves the best results. Statistical testing confirms that its predictions outperform those of the other models significantly.