Abstract: The traditional detection method of the protein content of wheat was tedious and time-consuming. NIRS (Near Infrared Reflectance Spectroscopy) and SPA-RBF artificial neural network were used to non-destructively measure the protein content of wheat in this paper. A representative set of correction samples was selected by SPXY algorithm, and then the spectral data was pretreated with first derivative and SNV methods to enhance spectral features, on the basis of which, making use of SPA to extract sensitive wave points which are used to establish SPA-RBF neural network model of wheat grain protein. Root-Mean-Square Error of Prediction (RMSEP) and prediction correlation coefficient (R) were 0.26576 and 0.975 respectively, which could basically complete the division that was used in grain reserves and food processing profession and breeding preliminary generation. The study showed that: NIRS combing with SPA-RBF neural network could achieve the detection of the protein content of wheat, which could satisfy the need of non-destructive and real-time detection of wheat to meet the development of modern agriculture.