Abstract: To study the feasibility of multi- spectral imaging technology for rapid and non- destructive identification of wheat varieties, multi-spectral images that covered the range of 405-970 nm from 500 samples of 5 wheat varieties were collected to find out the specified spectral, color and morphological characteristics by using VideometerLab multi-spectral image acquisition equipment. The 5 wheat varieties were qualitatively identified by principal component analysis. Likewise, the recognition accuracy of 3 different models (neural network, support vector machine and random forest) was compared based on spectral features and spectral image features. The results showed that BPNN method had the best performance, which was 100% and 91.25% for the modeling set and prediction set respectively, when 19 spectral eigenvalues were employed. Moreover, the satisfactory recognition accuracy, which was 100% and 98.4% for the modeling set and prediction set respectively, was also achieved when 19 spectral features and 6 image features were integrated. It suggested that multi-spectral feature fusion based on BPNN can effectively improve the recognition accuracy of wheat varieties and provide a new way for rapid nondestructive detection of wheat varieties.