The wavelength selection is an important step in the spectra modeling analysis. In the present paper, three wavelength selection methods, including correlation coefficient (CC), moving window partial least squares (MWPLS) and uninformative variables elimination (UVE), were studied for the determination of oil yield in oil shale using near-infrared (NIR) diffuse reflection spectroscopy. The above methods were used to eliminate the redundant and irrelevant variables in spectral data for enhancing the analytic efficiency and predictive ability of calibration model. The effects of thresholds of CC, window width of MWPLS and noise matrix of UVE were studied. Partial least squares regression was used to build prediction model for predicting oil yield in oil shale, and the performance of PLS models constructed with and without the using of wavelength selection methods were compared. The results show that any of the three methods can simplify the calibration model and improve the performance of model. By using UVE, the total number of wavelength variables of spectral data, the RMSECV of calibration model and the RMSEP of prediction model were decreased by 22.8%, 9.3% and 4.5%, respectively.