Mixed pixel in hyperspectral image is actually nonlinear mixing of endmembers,which is caused by multiple reflectances and scattering.The traditional endmember extraction algorithms based on linear spectral mixture model perform poorly in finding the correct endmembers.Considering the physical characters of hyperspectral imagery,a new method is proposed to introduce image Euclidean distance into Laplacian Eigenmaps for nonlinear dimension reduction.The proposed method can discard efficiently the redundant information from both the spectral and spatial dimensions．Endmembers are extracted by looking for the largest simplex volume from low-dimensional space.Experimental results demonstrate that the proposed method outperforms the PCA and Laplacian Eigenmaps algorithm.