Postharvest classification of banana (Musa acuminata) using tier-based machine learning

•A practical noninvasive classification of Musa acuminata banana fruit based on tier.•Machine learning classifies postharvest banana from its color and size features.•Length size of banana tier’s top middle finger, a good predictor for banana class.•Random forest classifier achieves 94.2% accuracy, better than other classifiers.•Reject banana class has low predictability given only its color and size features. Manual classification of horticultural products contributes to postharvest losses but technology and emerging algorithms offer solutions to reduce such losses. A practical fruit classification of banana (Musa acuminata AA Group 'Lakatan') using machine learning is developed based on tier-based classification instead of classifying individually (“finger”) for practical purpose. Fruit were classified into extra class, class I, class II and reject class, and compared using three widely-used machine learning classifiers – artificial neural network, support vector machines and random forest. Given only four features of banana tier, the red, green, blue (RGB) color values and the length size of the top middle finger of the banana tier, all three models performed satisfactorily. The highest classification accuracy of 94.2% was achieved using random forest classifier. In addition, ignoring the reject class, which cannot be easily predicted using only the given features, at least 97% accuracy can be achieved in all other three classes. Non-invasive tier-based classification is a practical postharvest technique that can be applied not only for banana but also for other fruit and horticultural products.