In this paper a road network extraction algorithm for suburban areas is presented. The algorithm uses colour infrared (CIR) images and digital surface models (DSM). The CIR data allow a good separation between vegetation and roads. The image is first segmented in two steps: an initial segmentation using the normalized cuts algorithm and a subsequent grouping of the segments. Road parts are extracted from the segments and then first connected locally to form subgraphs, because roads are often not extracted as a whole due to disturbances in their appearance. Subgraphs can contain several branches, which are resolved by a subsequent optimisation. The optimisation uses criteria describing the relations between the road parts as well as context objects such as trees, vehicles and buildings. The resulting road strings, represented by their centre lines, are then connected to a road network by searching for junctions at the ends of the roads. Small isolated roads are eliminated because they are likely to be false extractions. Results are presented for three image subsets coming from two different data sets, and a quantitative analysis of the completeness and correctness is shown from nine image subsets from the two data sets. The results show that the approach is suitable for the extraction of roads in suburban areas from aerial images.