Abstract: Soybean mildew has posed a significant threat to the economic viability and food safety in agricultural storage. Mildew detection is also limited to the technical problem during storage. Conventional detection, such as visual inspection, sensory evaluation, and chemical analysis, can often suffer from some challenges, such as time-consuming procedures, subjective judgment errors, and destructive sampling. It is often required for the timely and accurate identification of mildew. Therefore, it is important for the more advanced, efficient, and non-invasive detection. In this study, the Low-Field Nuclear Magnetic Resonance (LF-NMR) was adopted to achieve the rapid, non-destructive detection and grade discrimination of the internal water state and mildew degree of soybeans. Specifically, the soybean samples were stored at a temperature of 40 ℃. While a relative humidity of 75% RH for 21 days was selected for the test objects. These specific environmental parameters were chosen to investigate the mildew development. The mildew was more prone to occur due to the high-risk storage scenarios. Furthermore, the LF-NMR spectroscopy was carried out on the 0th, 7th, 14th, and 21st days. In addition, the LF-NMR instrument was calibrated before each detection, according to the standard samples with the known water content for the high accuracy and reliability of the measurement. The soybean samples were uniformly placed in the cylindrical glass tubes with a diameter of 10 mm, and the tubes were then inserted into the sample holder of the LF-NMR instrument. The parameters were optimized, including the magnetic field strength, radiofrequency pulse sequence, and acquisition time, using preliminary experiments to obtain the high-quality relaxation spectra. According to the spectral characteristic parameters, three discrimination models were obtained to determine the grade of the soybean mildew, namely Support Vector Regression (SVR), Random Forest (RF), and Gradient Boosting Machines (GBM). In the construction of the RF model, the systematic adjustment was performed on the number of decision trees, the number of features for splitting at each node, as well as the minimum number of samples required to split an internal node. The GBM model was fine-tuned to optimize the parameters, such as the learning rate, the number of boosting iterations, and the maximum depth of the trees. The results show that there were three water phases in the relaxation spectra of soybeans at each time point: bound water (0.1 ms ＜ T₂₁ ＜ 10 ms), semi-bound water (10 ms ＜ T₂₂ ＜ 150 ms), and free water (150 ms ＜ T₂₃ ＜ 1000 ms), with the signal amplitudes and water proportions of A₂₁, A₂₂, A₂₃ and P₂₁, P₂₂, P₂₃, respectively. The water content of each phase changed regularly during mildew. At the initial stage of mildew, the metabolic activities of the mold microorganisms triggered the breakdown of the cell walls and membranes in soybeans, thus causing the release of the bound and semi-bound water into the free state. There was an increase in the free water proportion P₂₃. As the mildew progressed, the free water was gradually consumed by the growing mold for its physiological activities, leading to a subsequent decrease in P₂₃. Meanwhile, the continuous rise of the semi-bound water proportion P₂₂ was attributed to the formation of new hydrogen bonds or other interactions between water molecules and the macromolecules in soybeans during mildew. Correlation analysis showed that the A₂₃ and P₂₃ shared high sensitivity and reliability. There were some changes in the water state during mildew of the samples. Thus, some influencing factors were then determined as the key input parameters for modeling. A comparison was made on the confusion matrices and multiple indicators of performance, including the accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve. It was found that the SVR model performed the best in terms of generalization and classification accuracy, compared with the RF and GBM models. The SVR model was also utilized to accurately classify the different mildew levels, according to the non-linear relationships between the input spectral parameters and mildew grades. The learning curve analysis further confirmed that the SVR model performed the best in terms of reliability and fitting in the classification of the mildew grades. The higher prediction accuracy and better stability were achieved, indicating its strong adaptability to various soybean samples. Efficient and accurate technology was obtained to detect soybean mildew for the intelligent storage of agricultural products. In practical applications, this LF-NMR-based detection can be integrated into the large-scale storage facilities of soybeans, enabling real-time monitoring of mildew status. Early warnings of potential mildew can greatly contribute to taking timely measures, such as ventilation, humidity, or separating contaminated batches, thereby minimizing economic losses. Moreover, the findings can also provide valuable references for similar detection for agricultural products, thereby contributing to food safety and quality control in the agricultural industry.