Near-infrared diffusion reflectance spectroscopy is a fast technique that can provide component information about intact soybean samples. We have combined this technique with partial least-squares (PLS) regression to perform a quantitative determination of protein and fat contents in soybean samples. In calibration set, the NIR model determination coefficient R2 of protein and fat is 0.9930 and 0.9752 respectively, and the relative standard deviation (RSD) is 0.76% and 1.3% respectively. The correlation coefficient r of validation set is 0.9473 and 0.8695 respectively. This NIR model is used to predict the contents of protein and fat in 264 soybean samples, using R-error to assess the deviation of analysis results. The minimum RSD of prediction of protein and fat is 0.04% and 2.46% respectively, and the maximum RSD of prediction of protein and fat is 2.45% and 4.25% respectively. These results are of great importance in early screening of crop breeding.