To quickly and accurately classify the maturityof blueberries, this study established a discriminant model for blueberry maturity based on near-infrared spectroscopy detection technology and deep forest algorithms. A LabSpec 5000 spectrometer was used to collect three different maturity levels of blueberry standard samples, and a total of 150 spectral samples were obtained. To determine the optimal number of input model features, the original spectral data was subjected to SavitzkyGolay convolution smoothing, and then principal component analysis was used to reduce the smoothed data to 4 principal components. The polynomial feature derivation method derived 2nd, 3rd, 4th, and 5th order features for each principal component. The optimal feature derivation order in the deep forest was considered 4th order. To test the maturity discrimination effect of the deep forest, it was compared with random forest, extreme gradient boosting tree algorithm (xgboost), and stacking fusion model. In the comparison, the optimal hyperparameter combination for each model was determined. The deep forest and stacking fusion model used manual parameter tuning, while random forest and xgboost used a Bayesian optimization algorithm for hyperparameter optimization. The model evaluation indicators were accuracy, confusion matrix, receiver operating characteristic (ROC) curve, AUC measurement, and anti-noise ability. The results showed that on the test set, the accuracy of the deep forest and stacking fusion model was 95.56%, while that of random forest and xgboost was 93.33%. The AUC value of deep forest was 1, while that of random forest, stacking fusion model, and boost were 0.99, 0.98 and 0.96, respectively. The anti-noise ability of deep forest and stacking fusion model was better than that of random forest and xgboost. Overall, the deep forest model in this study had a better discrimination effect than the other three models and provided technical support for blueberry maturity discrimination.