Monitoring the states of tobacco leaves during the flue-curing process is crucial for automating the tobacco industry. While much of the existing research on tobacco leaves state recognition focuses on the temporal state of the leaves, the morphological states are often neglected. Moreover, these studies typically use a limited number of non-industrial images for training, creating a significant disparity with the images encountered in real-world applications. To resolve these issues and investigate the potential of deep learning algorithms in identifying the morphological states of tobacco leaves in real industrial scenarios, this work has first developed a comprehensive, large-scale dataset. This dataset focuses on the states of tobacco leaves in real-world flue-curing houses, specifically recognizing the degrees of yellowing, browning, and drying. Then, a deep learning benchmark for this dataset using multiple deep learning networks is established. Furthermore, an efficient deep learning method is proposed to enhance the predictive performance of the deep backbone network. This is achieved by integrating the spectral characteristics of tobacco leaves images and filtering out color noise. The experimental results demonstrate the effectiveness of our approach in this task. The prediction accuracies achieved for the yellowing degree, browning degree, and drying degree are 83%, 90.5%, and 75.6% respectively. The overall average accuracy is 83%, satisfying the requirements of practical application scenarios.