Image-Based Classification of Golden Melon Ripeness Using Convolutional Neural Networks and Data Augmentation

Abstract

Accurate assessment of fruit maturity is essential for postharvest quality control and determining optimal harvest timing. This study evaluates machine learning methods for classifying the maturity levels of Golden Apollo melons using image data. A self-developed dataset consisting of 230 images was collected and categorized into four maturity stages: 47, 53, 60, and 67 days after planting (DAP). Four modeling approaches were tested: PCA with Support Vector Machines (PCA + SVM), PCA with Neural Networks (PCA + NNs), a Convolutional Neural Network (CNN), and a CNN with data augmentation. The augmentation process included random rotations (≤20°), horizontal and vertical shifts (≤10%), zooming (≤20%), and horizontal flipping, applied in real time during training. Model performance was assessed using precision, recall, F1-score, and accuracy. The augmented CNN achieved the highest accuracy of 86%, outperforming the other models. Confusion matrix analysis showed that the 60 DAP class was the most challenging to classify due to its similarity to adjacent ripening stages. The contribution of this research includes both the creation of a labeled image dataset and the demonstration of deep learning superiority for agricultural maturity classification.