In this paper, we use a morphological segmentation method called watershed for segmenting roof of 'orthophotoplan' images. This work takes place in a global approach which consists in recognizing a roof of aerial images among a knowledge database and bending out 3D models automatically generated from geographical data. The main aim of this work consists in defining the best couple of colorimetric invari- ant/gradient (among 24 colorimetric invariants and 14 gradients tested) used as input of watershed algorithm in order to obtain the best segmen- tation of roof. The tests are made on a database of 67 roofs containing a certain heterogeneity (illumination changes, shadows, etc) and evalu- ated with the Vinet criteria (including a ground truth image) in order to prove the robustness of the proposed strategy.