Abstract: Informative wavelengths were selected from the near-infrared reflectance spectroscopy (NIRS) of maize by genetic algorithm and partial least squares regression (PLS). A calibration model for determination of starch content was built by PLS based on the selected wavelengths of NIRS. The result showed that the root mean square error of calibration (RMSEC), root mean square error of cross validation (RMSECV) and root mean square error of prediction (RMSEP) derived from the calibration model based on the 11 selected wavelengths were 0.30%, 0.35% and 0.27%, respectively. And the coefficients of relationship between measurements and predictions for calibration and independent validation datasets were 0.927 9 and 0.939 0, respectively. The accuracy of prediction was better than the model based on the full NIRS data. It was proved that modeling by PLS based on the feature selection with genetic algorithm and PLS was a simpler, effective and more accurate means for determination of starch content in maize.