Objective: To establish a rapid method for the determination of stem content in finished tobacco by near infrared spectroscopy. Methods: The finished tobacco mixed with different proportion of stem silk （5%, 8%, 11% and 14%） was used as the test object. The near infrared spectrum information of pure stem silk, pure leaf silk and tobacco mixed with different proportion of stem silk was collected, respectively. Orthogonal Partial Least Squares Discrimination Analysis （OPLS-DA） was used to identify the overall difference of the near infrared spectral information of the above three kinds of samples. Variable Importance for the Projection （VIP） feature spectrum and full spectrum were used to establish linear non-negative regression models for stem silk samples of different proportions, respectively, and the accuracy and stability of the prediction models were evaluated. Results: ① The OPLS-DA pattern recognition analysis showed that there was a great difference in the overall NIR information of the three kinds of samples. ② Based on the VIP value greater than 1, the spectral wave number range was 4 000~4 165, 4 582~4 609 and 4 612~5 264 cm^-1. Wavelength range with large variance: 4 000~6 100 and 6 900~7 500 cm^-1. ③ The accuracy and stability of the prediction model established by the variance spectrum screening wavelength variable were superior to VIP algorithm and superior to the full wavelength variable. Conclusion: The method of NIR combined with linear non-negative regression coefficient regression for the rapid determination of the content of stem in finished tobacco has good stability and accuracy.