For the purpose of applying the spatial observation technology into the urban geographical information system, a new method based on remote sensing images is proposed to detect and map urban roads automatically. It builds a hierarchical objects network to organize image structure, getting precise processing units. Then the unsupervised learning integrating salient features is performed to analyze the explicit and implicit information, and to label the road targets in remote sensing images efficiently. It also applies spatial smoothing which incorporates contextual information to eliminate the adverse effects caused by background disturbance, occlusion and so on. After vectorization procedure, the contours and relative parameters of roads can be given. Experiments demonstrate that the proposed method achieves high exactness and robustness when detecting and mapping manifold roads in most high-resolution remote sensing images, and it is valuable for the urban information system and digital city-system construction.