Automated driving is becoming a reality. In this new reality, High Definition (HD) Maps play an important role in path planning and vehicle localization. Lane boundary geometry is one of the key components of an HD Map. Such maps are typically created from ground level LiDAR and imagery data, which, while useful in many ways, have many limitations such as prohibitive cost, infrequent update, traffic occlusions, and incomplete coverage. In this paper, we propose a novel method to automatically extract lane boundaries from satellite imagery using pixel-wise segmentation and machine learning. We then convert these unstructured lines into a structured road model by using a hypothesis linking algorithm, which addresses the aforementioned limitations. We also publish a dataset consisting of satellite imagery and the corresponding lane boundaries for future authors to train, test, and evaluate algorithms.