Clustering and Optimum Band Selection of Hyperspectral Data Based on Particle Swarm Optimization

Nowadays, by developing the remotely-sensed systems and acquiring the hyperspectral data with the high spectral resolution, it is expected that phenomena have been accurately recognized, and as a result, clustering accuracy of corresponding land-cover classes have been increased. Despite this useful characteristic, spectral information with high dimension is a challenging problem in data processing and analyzing. That is why, by increasing the dimension of data, firstly clustering accuracy is improved to a threshold, and then it decreases. So in order to make an appropriate clustering, it is needed to choose the optimum bands of these images simultaneously or in other words, the two problems of clustering and choosing optimum bands should be considered together. In this paper a new method has been presented for clustering of hyperspectral data by PSO, in which both optimum centroids and the number of optimum bands of these images are achieved. High capability of PSO algorithm in global searching of the space problem, and at the same time, solving the multi-objective problems caused to use this algorithm for clustering the hyperspectral data. In this study, after surveying the concept of PSO-based clustering, the proposed method is presented based on solving clustering and choosing optimum bands simultaneously. Comparison of the obtained results for clustering of a scene of AVIRIS image has showed that clustering by PSO and proposed method with the accuracy of 74.22 and 76.57 respectively for Kappa coefficient gives better results than the other algorithms which have been used in this research.