Based on near infrared spectroscopy, seventy samples of wild medicinal plants of paris polyphylla from Guizhou, Guangxi and Yunnan Provinces were collected to identify their geographical origins. Multiplication signal correction (MSC), standard normal variate (SNV), first derivative (FD), second derivative (SD), savitzky-Golay filter (SG), and Norris derivative filter (ND) were conducted to optimize the original spectra of fifty samples of training set. The results showed that the method MSC combined with SD and ND presented the best results of spectra pretreatment. According to spectrum standard deviation, spectrum range (7 450～4 050 cm-1) was chosen and principal component analysis-mahalanobis distance (PCA-MD) method was used to build the model. Its first three principal components, i.e. cumulative contribution, determination coefficient (R2), root-mean-square error of calibration (RMSEC) and root-mean-square error of prediction (RMSEP) were 89.44%, 97.58%, 0.179 6 and 0.266 4, respectively, and the prediction accuracy is 90%. Furthermore, according to variable importance plot (VIP), spectrum range (7 135.33～4 007.35 cm-1) was chosen and partial least square discrimination analysis (PLS-DA) was applied to establish the model. Its first three principal components cumulative contribution, R2, RMSEC and RMSEP were 89.28%, 95.88%, 0.234 8 and 0.348 2, respectively, and the prediction accuracy is 100%. Comparing the two methods, we found that spectrum range chosen by VIP and model built by PLS-DA could provide greater accuracy in identifying paris polyphylla from different origin areas. The method supplied foundation for authenticity and quality evaluation of traditional Chinese medicine.