In this thesis we address the problem of semantic segmentation in geospatial data. We investigate different deep neural network architectures and present a complete pipeline for extracting road network vector data from satellite RGB orthophotos of urban areas.

Firstly, we present a network based on the SegNeXt architecture for the semantic segmentation of the roads. A novel loss function is introduced for training the network. The results show that the proposed network produces on average better results than other state-of-the-art semantic segmentation techniques. Secondly, we propose a fast post-processing technique for vectorizing the rasterized segmentation result, removing erroneous lines, and refining the road network. The result is a set of vectors representing the road network. We have extensively tested the proposed pipeline and provide quantitative comparisons with other state-of-the-art based on a number of known metrics. This work has been published and presented at the 14 th International Symposium on Visual Computing, 2019.

Finally, we present an altogether different approach to road extraction. We reformulate the task of extracting vectorized road networks as a deep reinforcement learning problem with partially observable state-space and present our preliminary results and future work.