Deep learning approach for automated ‘Kent’ mango maturity grading in compliance with Peruvian standards

Abstract

Deep learning, particularly convolutional neural networks (CNNs), has significantly advanced automated fruit classification based on image analysis. However, accurate classification of Mangifera indica L. remains challenging due to high variability in external appearance and the subjectivity of visual maturity assessment. Misclassification contributes to post-harvest losses, reduced market value, and inconsistencies in quality control. This study develops a CNN-based model for classifying 'Kent' mangoes according to the Peruvian Technical Standard (NTP) 011.025:2023. A dataset of 603 labelled images was used to optimise the CNN architecture, systematically evaluating convolutional and pooling layers, image resolution, and training cycles. The optimised model, trained on 32× 32 pixel images, achieved 96.04 % classification accuracy, 90.91 % recall, and an F1-score of 93.57 %. To validate model robustness, 5-fold cross-validation demonstrated minimal accuracy variation (±0.5 %), while external evaluation achieved 95.8 % accuracy, confirming its real-world applicability. The lightweight single-layer CNN ensures scalable, low-cost implementation for automated sorting systems, reducing computational demands while enhancing classification efficiency. These findings establish deep learning as a viable and cost-effective solution for post-harvest fruit classification, ensuring greater consistency in quality control and supporting sustainable agricultural practices.