In order to improve the accuracy of spectroscopy analysis and reduce the modeling wavelength numbers, empirical mode decomposition (EMD) and successive projections algorithm (SPA) were applied together in the measurement of benzoyl peroxide (BPO) additive amount in flour by near infrared spectroscopy. Spectra of flour samples into which BPO were added were collected. Firstly, EMD was implemented to eliminate the noise of original spectra, and then SPA was employed to select the characteristic wavelengths. The precision of the model based on the processed spectra by EMD was greatly improved compared with the model based on the original spectra, with the calibration determination coefficient R2cal increased from 0.81 to 0.899 and the prediction determination coefficient R2pred increased from 0.755 4 to 0.86. Seven characteristic variables were selected from 512 wavelength variables by SPA. And the performance of the model built by the selected characteristic variables (R2cal is 0.863, R2pred is 0.86) was as good as full-spectrum model’s, while the number of modeling variables was greatly reduced by 96.4%. The results indicated that empirical mode decomposition and successive projections algorithm can be effectively used to denoise the spectra and selected characteristic wavelengths for the detection of BPO addition in flour. The seven selected wavelengths in this paper can be a reference for designing portable BPO detection meter.