A Transfer Learning Method for the Protection of Geographical Indication in China Using an Electronic Nose for the Identification of Xihu Longjing Tea

Abstract

Xihu Longjing (XHLJ) tea is one of the most precious teas in China; however, illegal dealers mislabel lower-quality teas as XHLJ tea and introduce them to the market. A rapid, easy, and accurate approach for the detection of the producing area and harvest time is necessary to protect the geographical indication of XHLJ tea as a Chinese product and to maintain market order. In this study, a MOS-based electronic nose (E-nose) was applied to distinguish different qualities of XHLJ tea. A transfer learning method called TrLightGBM (a combination of a transfer component analysis (TCA) strategy and a LightGBM classifier) is proposed to solve the problem of the differences in the distributions of data measured by the E-nose over long intervals, and can improve the classification accuracy of the XHLJ tea. Experimental results demonstrate that the proposed TrLightGBM model achieved the best performance for the identification of different producing areas and harvest times as compared to five other machine learning models (the support vector machine, random forest, XGBoost, LightGBM, and backpropagation neural network). The proposed TrLightGBM model can avoid the data re-collection and waste of the E-nose, and is a potential tool to maintain order in the Chinese tea market and promote the practical application of E-nose devices.