The combined use of available spectral and spatial information for object detection, which has been promoted by the advent of high spatial resolution hyperspectral imaging devices, now seems essential for many application domains (characterization of urban areas, agriculture, etc.). The proposed approach called "butterfly" is focusing on this issue and realizes a spectral-spatial cooperation scheme to split images into spectrally homogeneous adjoining regions (segmentation). The main idea of the method is to extract spatial and spectral features simultaneously. For achieving this goal, it establishes some correspondences between the spatial and the spectral concepts, in order to run alternately in the two spaces. Thus, the notion of partition specific to the spatial space is associated with the notion of classes in the spectral space. In parallel, the concept of latent variable owing to the spectral space is associated with the notion of image plans in the spatial space. The proposed scheme is therefore to update the features specific to each space (i.e. partition, classes, latent variables and plans) by the knowledge of the features in the complementary space and this recursively. An implementation of this generic scheme using a split and merge strategy is given. Experimental results are presented for a synthetic image and two real hyperspectral images with different spatial resolution. Results on the set of real images are also compared to those obtained with conventional approaches.