Abstract: The analysis and prediction model of alfalfa protein content was established with 63 alfalfa feed samples of different sources by using near-infrared reflectance spectroscopy （NIRS） and modified partial least-squares （MPLS） regression. The results showed that the standard error of cross validation （SECV） was 0.433 5, 1 minus the variance ratio （1-VR） was 0.920 1, and the standard error in validation （SEP） was 0.264 0. The combined data suggested that the protein content of alfalfa could be exactly predicted by the developed near-infrared reflectance spectroscopy.