This study proposes a deep learning-based multispectral image classification framework for distinguishing genetically similar mutant white rice varieties. An automated grain segmentation pipeline and a two-stage Convolutional Neural Network (CNN) are designed to preserve morphological features and to exploit spectral information to improve classification. Experimental results show that preserving original grain morphology improves classification accuracy by more than 15%. While RGB-based models achieve an accuracy of 85.9%, incorporating selected spectral bands at 500 nm, 540 nm, and 640 nm increases accuracy to 88.7%. The proposed two-stage CNN further improves performance to 90.6%, outperforming single-stage models by approximately 5%. These results demonstrate the effectiveness of combining multispectral imaging and deep learning for automated rice variety classification.
CHAN et al. (Thu,) studied this question.