The advance in technology over the past years has made it possible to launch different types of satellites equipped with many sensors to the orbit of the earth. This has led to the availability of tons of raw data that requires extensive analysis in different domains. The remotely sensed images were analyzed and inspected for automatic road network generation through the past years. Deep convolutional neural networks have proven their capabilities, producing state-of-the-art results. This paper

more »

... presents a comprehensive overview along with an experimental comparison of four widely used deep learning architectures applied to the automatic road network generation problem. The results show the advantage of LinkNet in terms of the best parameter size and MANet scores in total model size while UNet++ produces the best mean intersection over union score.