Total electron content (TEC) representing the integrated electron density from satellites to the receivers is one of the important parameters which can monitor the condition of space weather and provide valuable information for navigation improvement (Coster et al., 2008). With the rapid development of the technologies of ionospheric observation, such as the GPS-based measurement (Jakowski et al., 2002) and groundbased measurement, a mass of TEC data are available. Benefit from the continuous observation of GPS receivers, researchers produce global maps of total electron content (TEC-map) at hours interval (Mannucci et al., 1998). To generate the reliable TEC global maps, the International GNSS Service Ionosphere Working Group is set up in 1998 (Feltens, 2003). After the raw data are measured by the GNSS ground network, several independent computation centers compute the distribution of TEC worldwide, and the analysis centers will evaluate these TEC maps. Finally, the Ionospheric Associate Combination Center combines the products above and generate the reliable VTEC map, Vertical Total Electron Content map (Hernández-Pajares et al., 2009). Since 1998, IGS has provided abundant continuous global VTEC maps for space physics research. However, challenges still occur due to the lack of ground receivers, especially over the oceans and Southern Hemisphere (Rao et al., 2019). It is hard for other institutes or communities except IGS to fill data in these areas exactly concerning the complexity of the process to generate the IGS TEC. Fortunately, the flourish of Deep Learning during these years may bring a breakthrough to this problem. Despite the traditional interpolation methods (Orús et al., 2005), the Generative Adversarial Networks (GAN) in Deep Learning may provide a new way to solve the problem of data lack. Owing to the improvement of convolutional neural networks (Chua et al., 1993; Krizhevsky et al., 2012) and the progress of optimization algorithms, the generative models such as GAN (Goodfellow et al., 2014) can characterize the structure of the pictures and figure out the possibility distribution of input data. Learning from the training set, GAN is able to train a discriminator and a generator to map a probability distribution which approaches the real picture's (Radford et al., 2015). Thus, GANs can generate the artificial samples almost same as those in the training set. Based