Roads are essential for the generation and/or updating of old maps and geographical information systems (GIS). This paper
presents a new approach for modeling the centerlines and widths of road networks from very high-resolution (VHR)
satellite imagery at 0.5 m resolution. The proposed approach includes four main steps: (1) density-oriented fuzzy c-means
(DOFCM) algorithm has been applied to separate road and non-road pixels; (2) morphological operators have been applied
to eliminate noises, fill holes and reduce inconsistencies along edges; (3) road centerlines have then been extracted using
morphological skeletons and simplified using Douglas–Peucker algorithm; (4) the width of each road segment has been
determined as the mean value of the obtained widths at each pixel along the centerline of that segment. Compared with
manually digitized reference data, the results showed that the proposed approach has outperformed the most commonly
used approaches, Definiens eCognition software and fully convolutional networks (FCNs), with higher correctness and
lower root mean square error (RMSE).