Classification of Rice Seedling Structures Using ResNet-50 with Data Augmentation and K-fold Cross Validation

This study evaluates the efficacy of a Convolutional Neural Network (CNN) with ResNet-50 architecture for classifying rice seedling structures into three categories: normal sprouts, abnormal sprouts, and dead seeds. A dataset of 400 images was collected from the Seed Science and Plant Breeding Laboratory, augmented to 1600 images using horizontal flip, pad-crop, and rotation techniques. The model was trained using the Adam optimizer (learning rate=0.001) over 10 epochs and validated with 10-fold cross-validation. Results demonstrated significant accuracy improvements; the baseline model achieved 82.50% accuracy, while the augmented model reached 93.75%. Incorporating k-fold cross-validation further enhanced performance to 92.25% (no augmentation) and 99.67% (with augmentation). These findings underscore the importance of data augmentation in reducing overfitting and improving generalization for image classification.