Interactive network-based navigation over large urban environments raises difficult problems due to the size and complexity of these scenes. In this paper, we present a clientserver system allowing navigation over 3D cities in real time. Due to a novel progressive and hierarchical representation of 3D models of densely built urban areas, only perceptible details for all the regions visible from a given viewpoint are progressively streamed to visualisation clients. Furthermore, efficient coding methods are used to compress the representation data allowing quick start-up of the interactive visualisation with a highly-detailed model. This is achieved through a set of dedicated algorithms allowing a very large city model to be structured into a multi-resolution representation. The method efficiently exploits the fact that most automated modelling techniques of urban scenes provides 2D\frac{1} {2} models (building footprint, height, altitude, ...). So as to efficiently and faithfully model complex buildings, a procedural representation for roofs and facades is proposed. Finally, we present an MPEG4 compatible implementation based on the introduction of new node types with the associated bitstream.