In this paper, we propose a strategy of regions merging for roof detection which is made on pre- segmentation results. It is based on a 2D modeling of the roof ridges and region features. The preliminary segmentation is obtained by the watershed algorithm with an optimal colorimetric invariant and color gradient. The choice of an appropriate couple invariant/gradient permits to limit illuminations changes (shadows, brightness, etc) present on several roofs and increases the segmentation results. The watershed algorithm offers satisfactory results but produces an over-segmentation due to many germs (ie. local minima). This effect is reduced by using an appropriate selection of germs but can also be improved with a post-treatment based on regions merging. The proposed merging criteria is based on the 2D modeling of roof ridges (number of segments modeling the common boundary between two regions candidates to the fusion) and on the region features (contrast on boundary of two common regions, average color of region). The proposed strategy is evaluated on 100 real roof images with the Vinet criteria using a ground truth in order to demonstrate the effectiveness and the reliability of the proposed approach .