<p>The cultivation of dragon fruit is expanding rapidly around the globe, but its vulnerability to diseases and difficulty in accurately assessing fruit maturity pose significant risks to crop yield and profitability. Although recent advances have explored deep learning for fruit classification, there remains a need for comprehensive, objective, and scalable solutions that can reliably assess both maturity and quality under real-world farming conditions. To address this gap, this study systematically evaluates a suite of advanced deep learning models, including Vision Transformer (ViT), VGG16, ResNet50, EfficientNet, Xception, and InceptionV3, on a diverse, field-collected image dataset of dragon fruits, applying robust preprocessing, augmentation, and transfer learning techniques to enhance model generalizability. Our results show that the ViT model achieved the highest performance for quality grading, with an accuracy of 96.45% and strong precision, recall, and F1-score, making it highly effective in distinguishing between fresh and defective fruits. For maturity detection, the VGG16 model, optimized with early stopping, attained an accuracy of 95.58%, with a precision of 0.95, and a recall of 0.9633, outperforming the other tested models in classifying fruits as mature or immature. These findings provide new benchmarks for automated dragon fruit grading, demonstrating the potential of deep learning to overcome the limitations of manual assessment by providing rapid, consistent, and accurate classification. Furthermore, the present work can be intensified through the development of a mobile application for in-field use, expansion of the system to other crops, and integration with Internet of Things (IoT) sensors to refine predictions based on environmental factors, which would significantly improve crop quality and streamline the farming process, reduce waste, and optimize harvest timing.</p>