Abstract: Near infrared spectroscopy combined with chemometrics methods has been used to detect the geographical origin of honey samples. The samples were divided into the training set and the test set by kennard-Stone algorithm. After being pre-treated with first derivative and autoscaling, the spectral data were compressed and de-noised using wavelet transform (WT). The radical basis function neural networks (RBFNN) and partial least squares-line discriminant analysis (PLS-LDA) were applied to develop classification models, respectively. The performances of different wavelet functions and decomposition levels were evaluated in relation to the total prediction accuracy for the test set. For apple honey samples, when wavelet function was db1 and decomposition level was 2, both WT-RBFNN model and WT-PLS-LDA model produced the largest total prediction accuracy of 96.2%. For rape honey samples, when wavelet function was db4 and decomposition level was 1, WT-RBFNN model made the largest total prediction accuracy of 85.7%; while when wavelet function was db9 and decomposition level was also 1, WT-PLS-LDA model got the largest total prediction accuracy of 90.5%; The results indicated that linear WT-PLS-LDA model was more suitable for geographical classification of honey samples than no-linear WT-RBFNN model. Near infrared spectroscopy technique have a potential for quickly detecting geographical classification of honey samples.