Automated fruit maturity grading using deep learning with feature fusion

Abstract

Abstract Fruits are valued for their nutritional benefits, providing essential carbohydrates, vitamins, and dietary fibre. However, assessing fruit ripeness remains challenging, as it is governed by complex physiological processes and environmental factors that are not always reflected in external appearance. This difficulty is exemplified by Nam Dok Mai Si Tong (NDMST) mangoes, which retain a largely uniform yellow skin during ripening, making visual maturity grading unreliable. To address this limitation, we propose AFMG-DLFF (Automated Fruit Maturity Grading using Deep Learning with Feature Fusion), a multimodal deep learning framework that integrates external RGB image features with internal biochemical attributes. Visual features are extracted using DenseNet201, Inception–ResNetV2, and EfficientNetV2, while intrinsic traits such as total soluble solids, titratable acidity, and BrimA are encoded via a dedicated neural network. These complementary feature spaces are fused and optimised using Glowworm Swarm Optimization (GSO) for hyperparameter tuning. The model is trained with an 80:20 train–test split and early-stopping-based validation, achieving a classification accuracy of 97.86% for NDMST mango maturity stages. The results demonstrate that AFMG-DLFF achieves strong performance relative to the evaluated deep-learning baselines and remains competitive with selected literature-reported fruit ripeness classification methods, while relying only on accessible RGB imaging and standard biochemical measurements. This highlights its potential as a practical, non-destructive, and cost-effective solution for automated fruit maturity grading in real-world supply chains.