On-Line Monitoring of the Tobacco Leaf Composition during Flue-Curing by Near-Infrared Spectroscopy and Deep Transfer Learning

Abstract

The yield, quality, and output value of tobacco leaves are strongly affected by the flue-curing conditions. To effectively control the flue-curing and guarantee quality, it is necessary to quantify the chemical composition of tobacco leaves and provide timely feedback during this process. Therefore, the practicability of on-line monitoring of moisture, starch, protein, and soluble sugars for tobacco leaves by near-infrared (NIR) spectroscopy and deep transfer learning was explored. The results showed that the use of an NIR spectrometer equipped with a fiber-optic probe with a deep learning algorithm accurately predicted the content of these components during the curing process. The convolutional neural networks model showed greater potential for on-line monitoring than partial least squares and support vector machines. Furthermore, a network-based deep transfer learning strategy was crafted to include seasonal and temperature variability to accurately predict samples from a new harvest season in a curing barn. The overall studies indicated the efficacy of NIR diffuse reflectance spectroscopy as a rapid and nondestructive method for on-line and simultaneous determination of moisture, starch, protein, and soluble sugars in the flue-curing process to assist in making decisions.