The reduction of hyperspectral data dimensionality and classification based on recursive subspace fusion

A new method called recursive subspace fusion for the reduction of hyperspectral data dimensionality and classification is proposed in this paper. The new method includes three steps. First, the correlation information entropy is calculated from different correlated bands, and based on which the whole data space is divided into subspaces. At the second step, each subspace is fused into an image by the wavelet-based fusion method. Then the fused images and remained bands are considered as a whole space and we process it recursively as in above steps, until some given condition is satisfied. Lastly, the space with reduced data dimensionality is classified by using the Maximum Likelihood Classifier. The computer simulations are conducted on the AVIRIS data for the new method and the classical PCA as well as current SPCT method. The dimensionality is reduced from 100 to 5 bands. The experimental results show that the proposed method not only reduces much more data dimensionality of the hyperspectral images, but also gets higher classification accuracy 95.20%, compared with PCA 87.3% and with SPCT 95.0%.