It is practicable for dimensionality reduction of hyperspectral scenes using manifold algorithm such as isometric mapping (ISOMAP) and local linear embedding (LLE). However the two classical manifold algorithm are not suitable for solving the large-scale hyperspectral scenes. We elaborate the problems encountered in applying ISOMAP and LLE to dimensionality reduction of large-scale hyperspectral scenes, then an improved algorithm called IISOMAP-LLE, which is based on incremental isometric mapping (IISOMAP) and LLE, is proposed to represent the nonlinear structure of hyperspectral imagery that linear algorithm minimum noise fraction (MNF) could not discover. At last we demonstrate two experiments using large-scale AVIRIS and OMIS-II hyperspectral scenes to illustrate the approach. Experimental results prove that the IISOMAP-LLE not only is much better than linear algorithm MNF but also can avoid superiority decline of separability compared with MNF that encounterd in enhanced isometric mapping.