<![CDATA[The manual grading of tobacco, a critical process determining its market value, is inherently subjective, labor-intensive, and prone to inconsistency. This study investigates the application of deep learning to automate the quality classification of Madura sliced tobacco, a high-value agricultural commodity. A novel, high-fidelity dataset was created, comprising 1,065 high-resolution images captured under standardized lighting and environmental conditions. An expert tobacco sorter with over five years of professional experience meticulously labeled these images into four distinct quality grades: Grade A (premium), Grade B (medium), Grade C (lower), and Grade X (waste). This research presents a rigorous comparative analysis of four pre-trained deep learning architectures MobileNetV3-Small, ResNet18, MobileViTV2, and EfficientNet-B0 fine-tuned for this specific classification task. Employing a 5-fold cross-validation methodology, the models were evaluated on their ability to accurately classify the tobacco grades. The experimental results reveal that the lightweight MobileNetV3-Small architecture achieved the highest mean test accuracy of 56.87%±2.71%. A detailed error analysis indicated that all models performed well on the majority classes but struggled significantly with the underrepresented minority classes, a challenge attributed to the dataset's severe class imbalance. This study validates the potential of lightweight deep learning models for automating tobacco classification, offering a promising pathway toward enhanced objectivity and efficiency in the agricultural industry. Furthermore, it establishes a crucial benchmark and identifies class imbalance as the primary obstacle to be addressed in future research for developing a field-deployable system.]]>
