In the field of online maps generation based on remote sensing images, the most common method is based on the fully supervised learning or unsupervised learning. To address the problems of the resource-consuming problem challenge and poor generation quality, this paper proposed a novel semi-supervised map generation model based on generative adversarial networks, which aims to transform remote sensing images to online maps directly by using only a few precisely matched data and a large amount of unpaired data. In addition, we designed a semi-supervised learning strategy based on transformation consistency regularization and data enhanced consistency, which overcomes the inconsistency problem caused by imprecise paired data and derives better generalization performance of the model. Adequate comparison experiments were conducted on different map datasets, which validate the semi-supervised method proposed in this paper as an effective and fast map generation tool.