In this study,one dimensional scale invariant feature transform(SIFT) algorithm was used to find stable characteristic wavelengths of the near infrared spectra(NIRS) of tobacco leaves.The reproducibility rate and reproducibility were calculated according to the wavelengths selected by SIFT based on the precision tests of spectra.The orthogonal table of L9(33) was adopted to optimize the parameters in the SIFT algorithm to make the reproducibility rate and reproducibility as high as possible.Based on the optimized parameters and the spectra of 10 representative samples tested on the master instrument,10 stable characteristic wavelength sets were selected,and the spectral responses of the union of these wavelength sets were used as the independent variables to build a NIRS model for predicting the total alkaloids in tobacco leaves by partial least squares(PLS) method,which was shortened as SIFT-PLS model.After the model was directly transferred to three slave instruments,the average relative prediction errors(MREs) of the total alkaloids in tobacco leave samples tested on the three slaves were all smaller than 6%,meeting the internal control requirements of tobacco enterprises.Nevertheless,when the whole wavelength model(WW-PLS) was directly transferred to the three slaves,MRE of only one slave instrument satisfied the requirements.Even the spectra of the samples tested on the three salves were corrected by piecewise direct standardization(PDS) method,MRE of samples tested on only one slave was lower than 6%.The robust SIFT-PLS near infrared spectroscopy model established by selecting stable characteristic wavelengths with SIFT method could be directly shared by the three slave instruments,without needing transfer sets,and correcting spectra of slave instruments or spectral calibration models.The method proposed in this paper could be used to realize the NIRS calibration model transfer without standards in the true sense.