Before being exported, mangoes generally undergo rigorous external and internal quality inspection processes in which near-infrared (NIR) spectral approaches are favorable for grading purposes. A successful NIR-based grading system depends largely on high-quality spectral sensors and the reliability of the classifier. Motivated by the high economic impact of Cat Hoa Loc mangoes (Mangifera indica L.), we demonstrated that the sweetness of that mangoes could be precisely graded based on a random forest (RF) classifier in a three-phase approach with a low-cost Visible-Near infrared (VIS-NIR) multispectral sensor chipset. This approach is so-called RPR because RF, Partial Least Squares regression, and RF were respectively applied to consecutively determine the significant VIS-NIR responses, the good features as input variables, and the reliable RF classifier via our formulated discriminant index (DI). The experimental results confirmed that higher classification accuracy was achieved by using the extracted latent features rather than the raw VIS-NIR data. The DI was effectively used as a reliability measure to select the optimal classifier among those of identical training and testing accuracies of 100% and 82.1%, respectively. Performance comparison between the optimal RF classifier with a Support Vector Machines classifier and a multinomial logistic regression showed that the developed RF classifier was superior in various performance indices. Therefore, it is promising to extend the proposed approach to more complicated fruit grading problems with sufficient VIS-NIR datasets that are acquired from low-cost multispectral sensors.