AI-powered classification of mangoes for variety recognition and maturity grading in export and premium markets

This is the authors' abstract. We don't add key points for this paper.

Thailand is a major producer and exporter of mangoes, but traditional grading based on visual inspection is often inconsistent. This study proposed an AI-driven two-stage framework for automated mango classification. The framework was developed and evaluated using a purpose-built image dataset of 1,432 mango samples, expanded to 2,864 images through augmentation, covering three commercial Thai varieties, Namdokmai Sithong, Mahachanok, and R2E2, and three maturity grades: immaturity, export maturity, and domestic maturity. In the first stage, mango variety was identified using a Random Forest classifier with optimized hyperparameters and morphological features, achieving 99.6% test accuracy and 98.7% cross-validated accuracy. In the second stage, variety-specific maturity models were applied, with InceptionV3 performing best for Mahachanok (80.0%), Gradient Boosting for Namdokmai Sithong (84.4%), and Random Forest for R2E2 (83.3%), with average cross-validated accuracies ranging from 78% to 79%. The developed models were implemented in a web-based application for rapid and nondestructive mango grading in modern trade, export, and premium markets. Using a publicly available Kaggle dataset, the proposed framework demonstrates the potential of AI to enhance grading consistency, reduce subjectivity, and support decision-making in postharvest quality management.