Aiming at remote sensing image data having properties of high-dimension and small amount of labeled samples

a dimensionality reduction algorithm called semi-supervised discriminative locality alignment based on graph is proposed.At first

a similarity graph and a penalty graph are constructed according to all labelled and unlabelled samples.Then

based on the principle that the dispersion between neighbours of the same class is minimum and that the dispersion between neighbours of different class is maximum

optimization goals on the similarity graph and on the penalty graph are respectively determined.At last

an optimal mapping from the high-dimensional space to a low-dimensional subspace can be obtained by simultaneously optimizing the two objective functions

which makes the dimensionality reduction of high-dimensional remote sensing images realized.Experimental results on ROSIS hyperspectral data show that the proposed algorithm can effectively improved the overall accuracy and Kappa coefficient of high-dimensional remote sensing images.