Rice is one of the main staples for almost half of the world’s population; therefore, accurate rice grain classification is an important task for quality control and assurance. However, manual rice grain classification is a laborious and error-prone task due to minute differences between rice types. This paper proposes a CNN transfer learning method for rice grain classification. The GrainSpace dataset of 79,200 rice grain images of eight types was assessed with five pre-trained architectures: ResNet50, EfficientNet-B0, DenseNet121, MobileNetV2, and VGG16. Among them, EfficientNet-B0 demonstrated the best accuracy of 99.51%, followed by MobileNetV2 with 99.06% accuracy and ResNet50 with accuracy of 99.04%. Gradient-weighted Class Activation Mapping (Grad-CAM) further enhanced interpretability by visualizing regions influencing predictions. This study clearly validates that transfer learning can solve the issue of accurate rice grain classification with reliability, scalability, and interpretability.