Addressing the training problem in cellular automata based hyperspectral image segmentation

Two important issues are still open within the field of hyperspectral image segmentation. On one hand, most methods usually perform an early projection of the hyperspectral information onto a less informative two dimensional representation. On the other hand, there is usually very little and dubious ground truth available, making it very hard to train and tune appropriate segmentation and classification strategies. This paper describes an approach to address these problems by considering the application of evolved cellular automata (CA) over the hyperspectral cube in order to produce homogeneous regions that allow to easily perform the segmentation task. This homogenization process is carried out without resorting to any form of projection while the CA is operating, thus preserving this way the spectral character ofthe information in the segmentation process. We also show that the evolution process we propose for obtaining the rules can be carried out over RGB images and the resulting automata used to process multidimensional hyperspectral images successfully, thus avoiding the problem of lack of appropriately labeled ground truth images. The procedure has been tested over synthetic and real hyperspectral images and the results are very competitive.