Hyperspectral Fruit and Vegetable Classification Using Convolutional Neural Networks with EfficientNetB3

The categorization of fruits and vegetables has gained significant importance in agriculture, food processing, and retail industries, driven by demands for automation, efficiency, and quality control. This study presents a high-performance deep learning method that employs EfficientNetB3, optimized through transfer learning on a hyperspectral image dataset comprising 4,320 images across 36 unique classes, obtained from Kaggle. The proposed model demonstrated outstanding performance, achieving 99.29% accuracy on the training set and 97.21% on the testing set, indicating robust generalization abilities. The decision to exclude data augmentation was made after initial trials revealed a decrease in model generalizability, consistent with findings from prior studies that emphasize similar constraints in hyperspectral domains. To facilitate practical application, we constructed the model utilizing Streamlit and OpenCV, allowing for real-time classification of fruits and vegetables for uses such as automated sorting, supply chain enhancement, and precision agriculture. A comprehensive evaluation methodology was employed, incorporating training and validation metrics, a confusion matrix, and an analysis of per-class performance based on precision, recall, and F1-score. A comparative adaptation study was conducted to evaluate the proposed model with over 20 benchmark deep learning architectures. While the custom CNN reached an impressive accuracy of 96.1%, it demonstrated limitations in generalization and robustness when compared to the EfficientNetB3-based model. This study confirms the efficacy of integrating hyperspectral photography with contemporary CNN architectures, providing a scalable and practical solution for advanced intelligent agricultural systems.