Improved generalization of spectral models associated with Vis-NIR spectroscopy for determining the moisture content of different tea leaves

Rapid nondestructive determination of the moisture content of tea leaves is the basis of intelligent control in tea processing. Visible and near infrared (Vis/NIR) spectroscopy can detect the fundamental vibrations of hydrogen group (O-H) in the organic molecules, and has become one of the most commonly used methods for rapid determination of the moisture content of tea leaves. However, the established spectral model often fails to predict new samples of different varieties or batches, and then it will limit the wide application of this technology. Hence, improvement in the model and removing the impact of these samples variation on the determination model become a key issue for quality control during tea processing. In this study, tea samples of nine batches from three varieties were adopted to verify the performance of model transfer in improving the generalization of spectral models, it is worth noting that the moisture content distribution of different varieties samples is obviously different. It can be found that spectral profile difference of the three batches of tea samples for each variety was effectively eliminated by model transfer. The prediction ability of the PLSR model was improved by correcting the spectra of the tested batches and varieties of tea leaves to solve the problem induced by sample differences. The determination model developed based on the variety of Longjing was successfully transferred to the other two varieties of Wuniuzao and Yingshuang, and the R-p(2), were improved from 0.4343 to 0.2066 to 0.7595 and 0.6376, respectively by using DS algorithm comparing with un-transferred models. This study demonstrated that the proposed model transfer strategy by correcting the spectral data could be a robust technique for the rapid detection of the moisture content in different types of tea leaves and lay a basis for industrial application.

Automated summary

The study focused on improving the generalization performance of spectral models and successfully eliminated the spectral profile differences of tea leaf samples from different varieties. The proposed model transfer strategy by correcting the spectral data proved to be a robust technique for the rapid detection of moisture content in different types of tea leaves.