In view of the time-consuming, labor-intensive, and costly problem of chemical detection methods in the national standard, the feasibility of near-infrared spectroscopy （NIRS） combined with chemometrics for rapid detection of rice protein was investigated. Based on strategies of variable selection, feature extraction and nonlinear modeling, BiPLS-PCA-SVM was constructed by combining reverse interval partial least squares （BiPLS） with principal component analysis （PCA） and support vector machine （SVM） to improve the performance of the protein regression model. In BiPLS-PCA-SVM, the optimal number of principal components （PCs） was selected by combining Monte Carlo cross validation with the predicted residual sum of squares, and the model parameters were optimized by genetic simulated annealing algorithm. To evaluate the performance of BiPLS-PCA-SVM, three different models, including Full-PLS, BiPLS and BiPLS-SVM, were established, and the prediction accuracy and model robustness of all models were systematically analyzed. The performance of BiPLS-PA-SVM model in predicting protein content was higher than that of other models, and the model established by using the optimal number of PCs and optimized SVM parameters had higher robustness and accuracy. For BiPLS-PCA-SVM, the determination coefficient, root-mean square error and residual predictive deviation of the validation set were 0.928 9, 0.196 7% and 4.024 6, respectively. The results showed that NIRS combined with BiPLS-PCA-SVM model could be used as a reliable alternative strategy to realize the rapid detection of protein content in rice.