Dependent component analysis for hyperspectral image classification
Qian Du, Ivica Kopriva · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2009
Independent component analysis (ICA) has been widely used for hyperspectral image classification in an unsupervised fashion. It is assumed that classes are statistically mutual independent. In practice, this assumption may not be true. In this paper, we apply dependent component analysis (DCA) to unsupervised classification, which does not require the class independency. The basic idea of our DCA approaches is to find a transform that can improve the class independency but leave the basis mixing matrix unchanged; thus, an original ICA method can be employed to the transformed data where classes are less statistically dependent. Linear transforms that possess such a required invariance property and generate less dependent sources include: high-pass filtering, innovation, and wavelet transforms. These three transforms correspond to three different DCA algorithms, which will be investigated in this paper. Preliminary results show that the DCA algorithms can slightly improve the classification accuracy.