Fourier transform near-infrared spectroscopy (FTNIR) spectra contain valuable information about the chemical constituents of tea. Furthermore, the chemical constituents and their content of tea reveal differences concerning different kinds of tea and, therefore, it is feasible to classify tea varieties by FTNIR. FTNIR spectra have the characteristics of high dimension, crests and troughs, spectral overlapping and staggering, so it is difficult to classify spectra. In order to solve this problem, possibilistic fuzzy discriminant c-means clustering (PFDCM) was proposed by introducing fuzzy linear discriminant analysis (FLDA) into possibilistic fuzzy c-means clustering (PFCM) for purpose of discriminating FTNIR spectra correctly. Interestingly, during fuzzy clustering FLDA can not only extract discriminant information from FTNIR spectra but can transform the data space. PFDCM can achieve the accurate classification of FTNIR spectra according to its fuzzy membership and typicality values, and it has some advantages such as fast speed and high accuracy. PFDCM is superior to fuzzy c-means (FCM) clustering in clustering spectra containing noisy data because the typicality values of PFDCM are no constraint that the sum of the membership degrees is one. Four varieties of tea samples, called Yuexi Cuilan, Lu’an Guapian, Shiji Maofeng and Huangshan Maofeng, were collected in this study, and a total of 260 tea samples were scanned over the range of 10 000~4 000 cm-1 by FTNIR spectrometer, and in the end the 1 557-dimensional data were acquired for further processing. For a start, spectral data were pretreated with multiplicative scatter correction (MSC) to reduce spectra scattering and noise effect and increase signal-to-noise ratio. Secondly, principal component analysis (PCA) was used to reduce the dimensionality of FTNIR spectra to seven. Thirdly, discriminant information was extracted from spectra and the dimensionality of data was transformed from seven to three by linear discriminant analysis (LDA). Finally, fuzzy c-means (FCM) clustering, PFCM and PFDCM were put into use, clustering data to classify tea variety correctly. The experimental results showed that under the condition of the weight index m=2.0 and η=2.0, the clustering accuracy rates of FCM, PFCM and PFDCM achieved 93.60%, 93.02% and 98.84%, respectively. After 25 iterations, FCM converged, but PFCM and PFDCM achieved 8 iterations and 23 iterations, respectively, and converged. As fuzzy clustering algorithms converged, FCM consumed the least time but the most time-consuming clustering was PFDCM. In conclusion, FTNIR coupled with MSC, PCA, LDA and PFDCM presented a classification model for the accurate identification of tea varieties.