Portable near infrared spectrometer combined with multi-class support vector machines was used to discriminate wolfberry fruit of different geographic regions. Data pre-processing methods were explored before modeling with the identification rate as indicator. To eliminate the influence of sample subset partitioning on model performance, multiple modeling and predicting were conducted and the statistical result of identification rate was utilized to assess model performance of different acquisition sites. The results showed that SVM model with raw spectra after pretreatment of second derivative and Savitzky-Golay filter smoothing showed the best predicative ability. And the model of every acquisition site except for site 5 exhibited good stability and prediction ability and its median and average of identification rate of external validation were all greater than 97%. It was suggested that surface NIR spectra of wolfberry fruit was applicable to accurate identification of geographic region, and portable near infrared spectrometer could act as an effective means of monitoring the quality of Chinese herbal medicine in circulation.