The calibration models for protein concentration in corn samples were established by Fourier transform near-infrared diffuse reflection spectroscopy and Partial Least Square (PLS) regression. According to the prediction effect the best wave band was chosen to provide a basis for designing a minitype near-infrared corn protein content analyzer. The spectra were processed by Multiplicative Scatter Correction(MSC) method firstly

then the original spectra

the first derivative spectra and the second derivative spectra were processed by Savitzky-Golay smoothing method. The following 5 wave bands

the whole region

combination region

the first overtone region

the second overtone region and the protein functional group bands were selected for establishing 15 calibration models adopting the original spectra

the first derivative spectra and the second derivative spectra respectively in each band. By adjusting the number of Savitzky-Golay smoothing points and number of PLS factors simultaneously and comparing with PLS computational experiments several times

the optimal number of smoothing points and number of factors for each models were obtained based on the prediction effect

and then the best one was selected. The research results show that the prediction effect using the first derivative spectra in the first overtone region is best one

and the correlation coefficient

the Root Mean Square Error of Prediction(RMSER) and the Relative Root Mean Square Error of Prediction(RRMSEP) for the corresponding model are 0.945

0.357

and 3.340%

respectively. It shows that the first overtone band (7 000~5 500 cm

－1

) can replace the whole band

and get better calibration effect.