Pixel classification into landscape categories is a common task since the development of remote sensing. Many approaches such as Markov field classification are often used, and enable to create urban atlases. However, these techniques are not adequate to analyse urban structures at the city scale, as they are local. This paper presents a post-processing method to transform labelled pixels into urban structures. The method identifies urban 'morphotypes' that are assembled into urban structures. Then the graphical language of 'chorems' is used to represent the structures graphically.