Unsupervised Pixel Clustering in Multispectral Images by Genetic Programming

In this paper an innovative approach to Spectral Pattern Recognition for multispectral images based on Genetic Programming is introduced. The problem is faced in terms of unsupervised pixel classification. Given an image consisting in B bands, the goal is to find the optimal number of clusters and the positions of their centres in the B-dimensional hyperspace, which allow the best possible description of the image. The pixels are then assigned to the clusters according to "minimum distance to means" principle. Furthermore the system is endowed with mechanisms able to avoid that cluster centres may be too close one another, which would favour an excessive increase in their number. As a result a goodquality clustered image is achieved. The output consists of the image divided into clusters, the proposed number of clusters, the centre coordinates and the spectral signature for any such cluster and solution fitness value. The results are compared against those achieved by another system, MultiSpec, which performs supervised classification, yet it is endowed with some features typical of an unsupervised classification system.