Dataset feature reduction using independent component analysis with contrast function of particle swarm optimization on hyperspectral image classification

Data dimensionality reduction is an important step in the preliminary image classification. Information quantity and resolution of hyperspectral images provide a chance to solve the problem better than multispectral images. In hyperspectral image classification, higher dimensionality of data could improve the capability of class detection as well as distinguish different classes with better accuracy. The method calculation of ICA is a transforming a random vector into another space which consists of independent components. Because marginal distribution is usually unknown, the possible solution is to reduce data dimension into an optimized contrast function to measure component independency. In this research, PSO algorithm is used to solve the optimization problem. PSO is used to distinguish the signal selected by two different contrast functions. The problem existed in gradient method is solved using PSO, that is getting trapped in local optimum. The result of feature reduction done by using ICA-PSO technique is then compared with the result of feature reduction done by using ICA algorithm and PCA. Furthermore, the result gained by using ICA-PSO is used to classify hyperspectral images. In this work, Support Vector Machine is used as classifier. Classification result obtained by using ICA-PSO dimensionality reduction on AVIRIS, the value of average accuracy (AA) is 0.8535, overall accuracy (OA) is 0.8310, and K is 0.785. Whereas on HYDICE, classification result obtained by using ICA-PSO dimensionality reduction is at 0.8783 for AA, 0.8625 for OA, K is 0.850.