Abstract：Hyperspectral data is nonlinear,which is caused by the existence of a variety of features and multiple reflectance in a pixel.Therefore that traditional linear reduction algorithms has little effect.According to the nonlinear structure of hyperspectral data,a manifold learning method is presented to find out the low-dimensional smooth manifold embedded in high-dimensional data space,so that nonlinear spectral dimensionality reduction can be carried out.Experimental results on simulated and real hyperspectral data show that machine learning dimensionality reduction methods based on isometric mapping(ISOMAP) or local tangent space alignment(LTSA) outperform principal component analysis(PCA) algorithm with the advantage of better endmember divisibility in hyperspectral images.

Keyword：hyperspectral imagery;nonlinear reduction;manifold learning;isometric mapping;local tangent space alignment;