Abstract: The oil content of soybeans can directly dominate oil extraction and breeding. This study aims to rapidly detect and then identify the oil content of soybeans. Excellent seeds were provided for the soybean oil extraction and breeding. A systematic investigation was also implemented to construct the optimal detection model for the soybean oil content. Different dimensional low-field nuclear magnetic resonance (LF-NMR) spectrum was used to detect the oil and water content of soybeans, according to the national standard. The magnetic resonance imaging combined with deep learning was then used to establish a detection model for the high and low oil content. Two-dimensional LF-NMR was introduced to qualitatively analyze the component overlap that failed to be distinguished in the one-dimensional spectrum. The experimental results showed that the LF-NMR oil and water content software performed best to more rapidly and accurately detect the oil content of soybeans. The T1-T2 2D-NMR spectrum was used to successfully treat the overlapping of the signals of not easily flowing water, free water and oil. It was an excellent tool to supplement one-dimensional descriptions, where the water and oil signals of soybean kernels were visually distinguished. At the same time, a novel analytical approach was used to offer new perspectives on soybean trade and breeding with great promise. The U-net++ deep learning model was utilized to train the sagittal, coronal, cross-sectional, and three-sided mixed datasets of MRI imaging. Among them, the cross-sectional evaluation indexes were better compared with other datasets, where the mIoU in the semantic segmentation part was about 0.9058, with a global accuracy rate of 0.9980. The trained model was used to identify and segment the soybean MRI images, and then rapidly identify the oil content rate from the oil content level. The cross-sectional datasets were taken as the segmentation inputs for the magnetic resonance imaging, indicating the more reliable and accurate. The model, when combined with the data, allowed more information to be extracted from the MRI image than that commonly seen from subjective observations or simple signal accrual. Different datasets with magnetic resonance images of soybean cross sections were the most suitable for image segmentation. The experiment demonstrated that the LF-NMR and MRI system rapidly and non-destructively captured the information on soybean oil content, which provided a new idea and technical support to the high oil breeding of soybeans. This system can also integrate the previously fragmented detection with the LF-NMR and MRI technology, particularly for subjective observation. Deep learning can be combined to reduce the subjectivity with the scientific rigor at the same time. This LF-NMR application can be extended into the field of food and agriculture. The finding can also provide data support and theoretical guidance for the study of seed breeding, processing and storage. Meanwhile, the LF-NMR non-destructive testing can be expected to be widely promoted as a novel research tool.