Road Extraction from Medium Resolution Satellite Images using Deep Learning Techniques

Roads are one of the most important earth features in the satellite imagery. Its analysis plays a great role in urban/rural planning and its development. Several methodologies have been developed earlier towards this direction but most of them have used high-resolution satellite images. This research uses the freely available medium spatial resolution multi-spectral satellite images, i.e., Sentinel-2A from European Space Agency (ESA) and implemented & analyzed various state-of-art deep learning models, and architectures which are specifically tuned for road extraction, i.e., ResUnet, ResUnet-a, U-Net, Attention U-Net, LinkNet, & Dilated-LinkNet. It has been identified that the Attention U-Net based model has out-performed the other models in road extraction from medium spatial resolution satellite. Also, the work analyzed the optimum loss function suitable for road extraction in such resolution. The work can be utilized for urban road extraction for societal benefits.