A non-destructive method for discriminating varieties of grapes by visible and near reflection infrared spectroscopy (VIS-NIRS) was developed. The spectral data of three varieties of grape samples were clustered by principal component analysis (PCA). The results indicate that Heiti grape sample can be totally separated from the other two. Mainaizi and Mulage grape samples were discriminated based on back propagation-neural networks (BP-NN) model. The three hidden-layer BP-NN model was built with the first ten PCs as inputs, and the dummy variety numbers of grapes as outputs. The correct answer rate 98.28 % of BP-NN model is achieved, which is better than the one achieved by the soft independent modeling of class analogy(SIMCA) method. Four effective wavelengths for variety discrimination are 453,493,542 and 668 nm. The correct answer rate of BP-NN model based on the spectra of effective wavelengths is 97.41 %.The result indicates that variety discrimination of grapes can be achieved rapidly and non-destructively by using VIS-NIRS with PCA and BP-NN.