The highways now offer more and more complex road junctions composed of many surrounding roads, overlapping each other with high curvature. Traditional road detection methods are not adequate for the rapid development of cities with increased complexity of the road junction shape. The major challenges in road extraction are varying spectral reflectance, lane markings, obstacles with different sizes, of various shapes, and intersected roads. However, very few researchers have attempted handling overlapped roads with low curvature only. In this study, a water flow-based semi-automatic approach is proposed for extracting road network with various shapes of junctions (Y-shaped with different acute angles), intersected and also for overlapped high curvilinear roads. Recognising the complex road junction is done with fewer automatically generated anchor points without human intervention, which detects the number of roads (branches) connected to that junction along the road's width, orientation and length with less computation time. Hence, from a manually selected seed point, the authors' algorithm can be automatically propagated throughout a whole road network with or without single lane or multiple lanes (lined, dotted or both). Experimental results show that this proposed approach can accurately and efficiently extract interconnected road network from QuickBird images with minimal seed points. © The Institution of Engineering and Technology.