Abstract: In order to improve the accuracy of near infrared (NIR) spectroscopy identification methods for tea at different altitude, the non-linear manifold dimensionality reduction methods of locally linear embedding (LLE) and laplacian eigenmaps (LE) were used to reduce the dimension of NIR spectral data, and compared with non-linear (KPCA) and linear (PCA) dimensional reduction methods. Discrimination models were established for tea at different altitude based on different dimensional reduction methods and least squares support vector machine (LSSVM) algorithm. Visualization of different dimensionality reduction results show that data processed by KPCA and PCA methods were more discrete. In particular, there were more overlaps between 400-800 m and 800-1 200 m samples. However, the same kind of sample points could be gathered well in three-dimensional space by the nonlinear manifold dimensionality reduction methods can. Tea at different altitude could be easily separated and the aggregation effect of the LE was better than that of the LLE. The results of models indicate the LE_LSSVM model had the best performance, with the prediction set accuracy and Kappa value of 100% and 1.00 respectively. Compared with PCA_LSSVM, KPCA_LSSVM and LLE_LSSVM models, the accuracy of prediction set was improved by 1.7%, 1.7%, 3.3% and Kappa values increased by 0.025, 0.03, and 0.05. The results show that LE and other nonlinear manifold dimensionality reduction methods were effective in reducing dimension of near infrared spectral data, simplifying model complexity, and improving model precision. The study provides a new means for rapid detecting for tea quality research.

Key words: tea, near infrared spectroscopy, nonlinear manifold dimensionality reduction methods, laplacian eigenmaps