The present paper proposed how to select characteristic near-infrared wavelength for soil total nitrogen by using successive projection algorithm (SPA). Spectral data are compressed by SPA in the first place to obtain the raw wavelengths. Then the group of wavelengths derived from SPA is screened by their contributions to the total nitrogen. The insensitive wavelengths for total nitrogen are eliminated, improving the parsimony of the calibration model. For the 85 soil samples in total nitrogen, SPA was used to select the raw wavelengths. After screening on contribution, the number of wavelengths dropped from 12 by direct SPA to 6. Finally, the calibration model using wavelengths selected by screening on contribution after SPA showed the correlation coefficient (R(p)) of 0.913 and the root mean square error of prediction (RMSEP) of 0.011%. This model is as precise as the one before screening on contribution, and more precise than the result derived from partial least square (PLS) for the whole spectrum. The results demonstrate that the number of wavelengths selected by SPA can be reduced without significantly compromising prediction performance using the screening on contribution. The 6 selected total nitrogen wavelengths in this paper can be a reference for designing smart filter NIR spectrometer.