Autonomous driving car is becoming more of a reality, as a key component,high-definition(HD) maps shows its value in both market place and industry. Even though HD maps generation from LiDAR or stereo/perspective imagery has achieved impressive success, its inherent defects cannot be ignored. In this paper, we proposal a novel method for Highway HD maps modeling using pixel-wise segmentation on satellite imagery and formalized hypotheses linking, which is cheaper and faster than current HD maps modeling approaches from LiDAR point cloud and perspective view imagery, and let it becomes an ideal complementary of state of the art. We also manual code/label an HD road model dataset as ground truth, aligned with Bing tile image server, to train, test and evaluate our methodology. This dataset will be publish at same time to contribute research in HD maps modeling from aerial imagery.