Starch content in rice is one of the important parameters in characterizing the nutritional quality of rice, and the starch content of rice produced in saline soils under different fertilization conditions varies. In this study, Raman spectroscopy combined with three machine learning models, support vector machine (SVM), feedforward neural network, and k-nearest neighbor classification, was used to classify and evaluate the effect of different fertilizer treatments on rice. The collected rice spectral data were normalized before machine learning, then preprocessed with multiple scattering correction (MSC), standard normal variable, and Savitzky-Golay filtering algorithms to improve the quality and reliability of the data. The evaluation indexes such as the confusion matrix and the receiver operating characteristic curve comprehensively analyzed the model's performance. The research shows that the MSC preprocessing method significantly improves the classification accuracy and prediction ability in all three models, and the classification accuracy was close to 100%, while the overall performance of the SVM models after various preprocessing is the best among the three machine learning methods. The predictive coefficient of determination, predictive root mean square error, and predictive average relative error of the starch content detection model built by the SVM model after MSC preprocessing were 0.93, 0.04%, and 0.20%, respectively, which indicated that its prediction had high accuracy and low error. The results of this study used Raman spectroscopy to carry out the identification of different fertilization techniques and rice starch quality correlation characteristics, providing theoretical and experimental support for the rapid identification of rice quality.