Road Extraction and Routing from Satellite Imagery by Image Segmentation using Deep Learning

In this thesis, we address the challenging task of interpreting large-scale satellite imagery by developing an automated system for generating semantic road maps and road graphs with speed limit predictions to enable efficient routing. We explore various convolutional deep neural networks, such as ResNet34, ResNet50, SeResNetX50, and InceptionV3, and conduct extensive studies on hyperparameters and loss functions to optimize the road extraction process. Our pipeline includes image pre-processing algorithms to handle varying image qualities, a model for road segment prediction, and post-processing techniques for graph extraction while retaining geographic information. The results demonstrate the effectiveness of our approach, showcasing the importance of appropriate model selection and optimization. The integration of graph extraction and geographic information enhances the routing process. Overall, this research contributes valuable insights into road extraction and routing from satellite imagery using deep learning, laying the groundwork for future advancements in this field.