This study presents an accurate and interpretable approach for classifying five rice varieties: Arborio, Basmati, Ipsala, Jasmine, and Karacadag, using a dataset of 75,000 images. A deep learning (DL) framework was employed, combining Convolutional Neural Networks (CNNs) and Vision Transformer (ViT) models for feature extraction. Specifically, three pretrained CNN models: MobileNet Version 2 (MNetV2), Residual Network 18 (ResNet18), and EfficientNet Base Model 0 (ENetB0), along with three ViT models: Data Efficient Image Transformer (DeiT), Shifted Window Transformer (STrans), and Mobile VT (MobileViT), were used to extract robust visual features. The extracted features were classified using ensemble-based machine learning (ML) models: Extreme Gradient Boosting (XGB), Random Forest (RF), Categorical Boosting (CB), Adaptive Boosting (AB), and Light Gradient Boosting Machine (LGB). The combination of MNetV2 and XGB achieved the highest accuracy of 99.63 %, demonstrating the effectiveness of this pipeline. To ensure model interpretability, Explainable Artificial Intelligence (XAI) techniques, including Gradient Weighted Class Activation Mapping (GradCAM), and GradCAM++, were applied to visualize decision-relevant regions. This research provides a high-performance and explainable solution for automated rice classification suitable for agricultural automation and food quality control applications.