Semantic segmentation applied to aerial imagery allows the extraction of terrestrial objects such as roads, buildings and even vegetation. Having large, detailed datasets of navigable roads, is of paramount importance in several application fields; namely urban planning, automatic navigation, disaster management. To reach this goal, extracting all roads in a given territory area is the first step. This paper presents a modern method to semantically segment aerial images for a road network extraction. We employ an encoder-decoder architecture to approach the problem of disconnected road regions faced by some existing methods. Using an FCN approach, the localization information was combined to the semantic one, to enable the reconstruction of the road by the proposed model, while being consistent with following the spatial alignment. The method was implemented and evaluated on the public dataset Massassuchets Roads. Results appear to be in full agreement with the theorical predictions and a significant improvement in road connectivity over some previous works; the proposed network achieved a precision of 87.86% and a recall of 87.89%.