An Improved Independent Component Analysis Method for Unsupervised Classification of Hyperspectral Data

Abstract

To solve the problem that the first-order and second-order statistics may be inadequate for obtaining a complete representation of the data,a high-order statistics-based method,kurtosis-based independent component analysis(KICA),is introduced to implement unsupervised classification of hyperspectral data.Aimed at the purpose that kurtosis can be very sensitive to outliers such as noise,the improved KICA(IKICA) model is proposed in the work when kurtosis is used as optimization criterion for the ICA problem.To evaluate the performance of the proposed algorithm and its application capability in unsupervised classification,IKICA is compared with maximum likelihood-based ICA and negentropy-based ICA,and the synthesized and real hyperspectral data acquired by Object Modularization Imaging Spectrometer(OMIS) and Pushbroom Hyperspectral Imager(PHI) are used.The results show that convergence speed and robustness are enhanced obviously and anti-noise capability is improved in the authors' work.The application result has high precision of classification.