New dimensionality reduction algorithms for hyperspectral imagery based on manifold learning

A new neighborhood selection method was proposed based on the image patch distance and applied to the manifold learning. Thus, a new nonlinear methods for hyperspectral dimensionality reduction was obtained. Considering the physical characters of hyperspectral imagery, the proposed methods combined both spectral and spatial information and, thus, kept the original characters of dataset well with the less loss in the useful information and less distortion on the data structure. Compared with other dimensionality reduction methods for hyperspectral imagery, the proposed methods can reserve effectively the spatial relationships between observation pixels in hyperspectral imagery after transformation. Meanwhile, the proposed methods can discard efficiently the redundant information of original data sets along both spectral and spatial dimensions. Experimental results on real hyperspectral data demonstrate that the proposed methods have higher classification accuracy than the other methods when applied to the classification of hyperspectral imagery after dimensionality reduction.