Abstract: Both spatial and spectral information of Congou black tea blends from superior fifth-grade and sixth-grade raw Keemun black tea leaves winnowed through 6 and 8 pores per inch were acquired by hyperspectral imaging technology. Spectral features were selected by successive projections algorithm (SPA), and all the obtained hyperspectral images were analyzed by principal component analysis (PCA). According to the weight coefficient, five dominant wavelengths were extracted for the collection of textual features by gray level co-occurrence matrix (GLCM). Subsequently, partial least squares (PLS), least squares-support vector machine (LS-SVM) and back propagation-artificial neural network (BP-ANN) classification models for prediction of the optimal blend ratio of raw tea leaves were developed based on the spectral features, textural features and data fusion, respectively and they were comparatively evaluated. It was indicated that the discrimination accuracy of the LS-SVM model based on data infusion was up to 94.5%. Our results may provide a theoretical basis for the digitization and standardized production of tea blends.

Key words: tea, blends, hyperspectral imaging, data fusion, quantification