Improved independent component analysis applied to classification hyperspectral imagery

To avoid the disadvantage of getting into local optimum solution with general numerical computation methods in the general independent component analysis and the restriction of neuron activation functions of neural learning algorithm, an improved model of independent component analysis (ICA) based on genetic algorithm was proposed for the unsupervised classification of hyperspectral data. In the proposed algorithm, the maximizing non-Guassianity was used to measure the statistical independence of the components, and the forth-order cumulant kurtosis was adopted as fitness function in genetic algorithm. In the application, the global optimum solution can be obtained and the fine plant classification can be implemented without any prior information when the proposed algorithm is applied to the push-broom hyperspectral technique imager (PHI) data. Moreover, compared with the conventional unsupervised classification algorithm of hyperspectral data, the proposed algorithm is more applicable and can obtain better precision and accuracy.