To achieve rapid and non-destructive identification of early mold deterioration stages in rice during storage, changes in moisture content, fatty acid value (FAV), and aerobic plate count (APC) were monitored during the mildew process. Compositional and structural changes were also examined using near-infrared spectroscopy (NIR) and Fourier transform infrared spectroscopy (FTIR). Machine learning algorithms were used to construct discriminant models for early mold deterioration stages based on three strategies: single-spectrum analysis, data-level fusion, and feature-level fusion. The results indicated that the rice mold deterioration process exhibited distinct stage-specific characteristics. During the initial phase (day 0–29), moisture content rose rapidly, then reached a dynamic equilibrium, and subsequently increased again. FAV continued to increase, while APC remained at a low level, suggesting that biochemical degradation predominated in the early stages of deterioration. During the mold outbreak phase (day 30–33), moisture content increased significantly and subsequently declined; FAV reached a peak and then decreased due to microbial utilization, and APC exhibited exponential growth, indicating that microbial activity became the dominant factor driving quality deterioration. Following spectral preprocessing and feature extraction, the classification accuracy of the support vector classifier (SVC) model based on NIR spectra was as high as 96.8%, compared with 93.5% for that based on FTIR spectra. In contrast, the SVC model employing feature-level fusion achieved a classification accuracy of 100%. In summary, NIR and FTIR spectra exhibited complementary advantages for monitoring early mold deterioration in rice, enabling precise discrimination of mold deterioration stages and providing a feasible spectral analysis strategy for rapid, on-site detection of the early mildew process of rice.