Discrimination of black tea fermentation degrees based on data fusion strategy

Fermentation is a key processing step for the quality of black tea. Tea polyphenols (catechins) are generally oxidized by the polyphenol oxidase and peroxidase to form the theaflavins and thearubigins. In this research, the Yinghong NO.9 of Yingde black tea was collected by the kind of one bud and two leaves. The data was collected during different black tea fermentation time using a portable near-infrared spectrometer and a Charge-Coupled Device (CCD) camera. A discriminant model was established for the black tea fermentation degree using near-infrared spectra, images, and the data fusion of spectra and images. Specifically, 204 samples of black tea at different fermentation time were collected to acquire the near-infrared spectrum and images. The content of tea polyphenols was determined using an ultraviolet spectrophotometer with a detection wavelength of 765 nm. The catechins concentration was measured by High-Performance Liquid Chromatography (HPLC) at a flow rate of 1 mL/min and detection wavelength of 278 nm. The contents of tea polyphenols and catechins decreased rapidly in the early period, tending to be flat at 4-5 h, and continued to fall off after 5.5 h. According to the changes in the tea polyphenols and catechins, the fermentation degree of black tea was divided into three stages: insufficient, moderate, and excessive fermentation. Savitzky-Golay smoothing was adopted to process the rough burrs of the original spectrum that were caused by noise interference. Then, the Competitive Adaptive Reweighted Sampling (CARS) and Successive Projections Algorithm (SPA) were applied to reduce the data dimensionality of near-infrared spectral variables, where the feature wavelengths were selected. Meanwhile, nine color feature variables were extracted from the images after shadow removal. Pearson correlation analysis between chemical components and color variables was conducted to extract the feature variables. In addition, the Principal Component Analysis (PCA) was employed to reduce the data dimensionality for the distribution of black tea fermentation samples. The PCA of spectral and image data showed the similar three fermentation stages were not separated significantly, indicating that PCA cannot effectively discriminate the fermentation stage. Finally, the discrimination models were established using the near-infrared, image, and their data fusion through Linear Discriminant Analysis (LDA) and Support Vector Machine (SVM). The comparison of the model showed that the performance of nonlinear SVM models was better than that of LDA models under the same conditions, indicating the unbalanced process of black tea fermentation. Furthermore, a single sensor failed to discriminate the fermentation degree. There was less performance in the models using a single sensor, due mainly to the complex change of fermentation information. In general, the maximum accuracies were only 83.82% and 73.53% for the prediction set of the discrimination models using near-infrared spectra and images, respectively. The performance of the middle-level data fusion models was significantly improved, compared with the models founded on a single sensor, or the low-level data fusion. The reason was that the low-level date fusion brought the variables irrelevant to the black tea fermentation. Among them, better performance was achieved in the SVM discriminant model that was established by SPA extraction of spectral variables and Pearson correlation analysis extraction of image variables, with 97.06% and 95.59% accuracies of calibration and prediction set. Consequently, a rapid and nondestructive method can be used to evaluate the degree of black tea fermentation under the middle-level fusion strategy using near-infrared spectroscopy and computer vision. A theoretical foundation was laid to establish a grade model and discrimination of black tea fermentation degrees. The finding can provide an important basis for the detection and automation of black tea fermentation.