Road Network Extraction from Satellite Images: A Comparative Study

Road networks are used extensively on daily basis for many applications. However it is not an easy task. This has motivated research in road extraction and generation from satellite images and as a result numerous algorithms have been proposed. In this paper we have over-viewed the state of art algorithms and compared and evaluated three of the state of art deep networks: residual network, squeeze excitation next dimension residual network and dual path network. The three networks were tested and evaluated on publicly available datasets of six cities. The plain ResNet network proved to be the best in the Road Network extraction problem.