Present work proposed multi-step wavelength selection methods based on scale invariant feature transformance（SIFT） algorithm of extracting features of spectra images to build robust near infrared spectra（NIRS） models for predicting total reduced sugar（TRS） in tobacco leaves. So that the models can be directly shared by multiple NIRS instruments and applied in long period. The SIFT method was firstly applied to select characteristic spectral points of ,U,c, based on several representative spectra of primary samples. Then the spectral points with lower standard deviance of the sample spectra（SDSS） values were eliminated from ,U,c, to obtain important characteristic spectral points of ,U,ic,. The two-step wavelength selection method was named as SIFT-SDSS. Furthermore，the moisture-unsensitive（MUS） wavelengths were selected from ,U,ic ,to get important and stable characteristic spectral points of ,U,isc,. The three-step wavelength selection method was named as SIFT-SDSS-MUS. The partial least square regression（PLSR） models for predicting TRS were established based on different wavelength sets by using 80% of 292 primary tobacco leaf samples according to their TRS concentration ranges as calibration set，which were harvested in 2011-2013. The results indicate that when the TRS model based on the wavelengths selected by the SIFT-SDSS method was transferred to six secondary instruments to predict TRS in another 77 samples harvested in 2011-2013，the mean of absolute relative errors（MAREs） obtained from the spectra tested on the six units were all below 6%，satisfying the internal control requirement of tobacco enterprises. For each year samples of 2014-2020 tested on 5 NIRS instruments，and each year samples of 2014-2019 tested on another NIRS instrument，the model built by the SIFT-SDSS can provide MARE lower than 6%. While MAREs of TRS in samples harvested in many years and tested on the 7 NIRS instruments，which were predicted by full-wavelength model and the model based on the SIFT were higher than 6%. The two models cannot be applied on multiple instruments in long period. The TRS model built by the SIFT-SDSS-MUS method has the fewest variables，but its transferability and long-time service ability is a little inferior to the model built by the SIFT-SDSS method. In conclusion，the TRS model established by important characteristic wavelength selection method of SIFT-SDSS is most robust，has strong interpretability and fast computation speed. It can be directly shared among seven NIRS instruments and continuously service on 6 instruments at least six years. The work amount on maintenance and transfer of the TRS model is therefore greatly reduced.