Improving satellite image segmentation using evolutionary computation

Segmentation is the process of dividing an imageinto disjoint regions. It is the most important task in image processing where the success of the object recognition depends strongly on the efficiency of the segmentation process. The most popular and important segmentation methods are clustering such asFuzzy c-Means (FCM), Iterative Self-Organizing Data (ISODATA) and K-means. Clustering methods depends strongly on the selection of the initial spectral signatures which represents initial cluster centers. Normally, this is done either manually or randomly based on statisticaloperations.In either case the outcome is unpredictable and sometime inaccurate. In this paper an unsupervised method based on Multi-Objective Genetic Algorithm (MO-GA) for the selection of spectral signature from satellite images is implemented. The new method worksby maximizing the number of the selected pixels and bymaximizinghomogeneitythrough the minimizing of the difference between the pixels and their spectral signature. The objective is to create best cluster centers as an initial population for any segmentation technique. Experimental results are conducted usinghigh resolution SPOT V satellite imageandthe verification of the segmentation results is basedon a very high resolution satellite image of type Quickbird. The spectral signatures provided to K-means and Fuzzy c-meansby MO-GA process increased the speed of theclustering algorithmto approximately4 timesthe speed of the random based selection of signatures.At the same time MO-GA improved the accuracy of the results of clustering algorithmstomore than 10% compared to the random statistical cluster centers selection methods.