Rice mold can cause nutrient loss and produce toxic substances that reduce its quality and infect other normal rice. In order to reduce the loss of rice caused by mold, moldy rice needs to be separated promptly. Hyperspectral technology is fast and nondestructive, so an attempt was made to detect rice mold using hyperspectral technology. Germinated rice and moldy rice have similar spectral characteristics and are easily misidentified as moldy rice, which affects the subsequent detection of rice mold degree. Therefore, it is proposed to use hyperspectral techniques combined with various pre-processing and discrimination models to distinguish germinated rice from moldy rice and to discriminate rice with different mold degrees. Sound, sprouted, moldy and germinated moldy rice samples were modeled to differentiate and detect mild, moderate, heavy and completely moldy rice samples. The spectral images of sound, germinated, moldy and mildewed rice samples were acquired using a hyperspectral acquisition instrument to extract the spectra in the region of interest (ROI) of the acquired images, and the average reflectance of the spectra within the ROI was used as the spectral characteristics of the rice samples. Pretreatment of the extracted spectral data with SNV, Normalize and MSC. The KS algorithm is used to divide the samples evenly in a ratio of 1∶3, into a prediction set for validating the effect of the model and a modeling set for establishing the relationship between the spectra and the samples. The PLSR, SVM and RF models were developed respectively, and the prediction effect of each model was evaluated by the prediction set correctness of the three models, and the discriminative model with the best effect was selected. In detecting sound, germinated, moldy and germinated moldy rice, the optimal discriminatory model was obtained as a random forest (Baseline-RF) model after pre-treatment by the baseline correction method. The discriminatory accuracy of the prediction set of the Baseline-RF model was 100%. In detecting rice mold degree, a comparison of the prediction results of different models showed that the SNV-RF model showed the optimal discriminative effect with no misclassified samples in the prediction set. The characteristic wavelengths were extracted from the lengthy original spectra to simplify the model, and the SNV-RF model was established with the spectra under the characteristic wavelengths. The results showed that the characteristic wavelengths selected using the CARS algorithm had good discriminative ability, and the overall discriminative accuracy was 97.5%. The experimental results show that the hyperspectral technique combined with the CARS-SNV-RF model can quickly and accurately discriminate the degree of moldy rice, which provides a certain theoretical basis and experimental reference for the rapid discrimination of moldy rice and is of great significance for improving the quality of rice and reducing the waste of rice.