Abstract: The local modelling based on cluster analysis was applied to NIRS analysis of tea to promote the predictive testing precision in determination of its caffeine and amino acid contents.The characteristic factors of spectral data of tea sample were collected,and crisp partition of the samples was done by cluster analysis.Through differentiation of the sample distances and inter-class distances,number of target sample with single model was determined.After completion of analyzing the characteristic spectral bands and of selection of wave sections,15 samples were taken at random for cluster analysis.Results of local modelling by PLS were obtained as follows: the values of average relative deviation of predictive testing of caffeine and amino acids were decreased from 5.80% and 6.14% (before cluster analysis) to 2.75% and 2.44% (after cluster analysis) respectively,showing a remarkable improving of the precision of modelling predictive testing.