The extraction of urban patterns from very high spatial resolution optical images presents challenges related to the size, the accuracy and the complexity of the data. In order to efficiently carry out this task, a multiresolution hierarchical approach is proposed. It enables to progressively segment several images (of increasing resolutions) of a same scene, based on low level criteria. The process, based on binary partition trees, is partially performed in an interactive fashion, and then automatically completed. Experiments on urban images datasets provide encouraging results which may be further used for detection and classification purpose .