Harvested grain is generally stored in large bins. The quality and quantity of stored grain are important as seeds may be threatened by molds, sprouting, insects, or small animals. Therefore monitoring bin contents is critical. Electromagnetic imaging has recently been proposed as a method to monitor bin contents. For example, trained regression Machine Learning (ML) models are capable of extracting bulk parameters of the grain within the bin (i.e., grain height, cone angle, and bulk real-imaginary permittivity). Knowledge of the permittivity can be mapped to moisture and allows inference of potential stored losses and planning of mitigating action. Unfortunately, training these models requires a large labeled dataset. Obtaining such a dataset for experimental installations is unfeasible, and generating synthetic datasets is time-consuming. In this thesis, we investigate the application of Generative Adversarial Networks (GANs) to electromagnetic imaging for grain storage. GANs are designed to generate data of a particular type, and once trained, GANs can be used to generate data in an inexpensive way. Also, a recent GAN variant, the CycleGAN, has been proposed to transfer data from one domain to another. Such functionality is applicable to the electromagnetic imaging problem of calibration. We propose and answer two research questions. First, whether or not a simple 1DGAN can be used for data augmentation and second, whether or not a CycleGAN can be used for calibrating uncooperative imaging systems. Results show that a 1D-GAN is too simple to be used for data augmentation as transmitter-receiver ordering is not preserved. The 1D-GAN is demonstrated to produce data with similar distributions to real data, which suggests that with appropriate architectural modifications the idea is sound. Preliminary tests on faux-experimental data show that the CycleGAN has promised to serve as a calibration method, with the successful recovery of bulk parameters demonstrated to within acceptable tolerances on calibrated phaseless data. While promising, additional testing and development are required to make the concept sufficiently robust for practical applications.