A discriminant manifold learning approach for hyperspectral image dimension reduction was proposed. In order to overcome the high dimensional and high redundancy of remotely sensed earth observation images, a modified manifold learning algorithm was suggested for dataset linear dimensional reduction to improve the performance of image classification. The proposed method addressed the discriminative information of given training samples into the current manifold learning framework to learn an optimal subspace for subsequent classification, in particular, the linearization of discriminant manifold learning is introduced to deal with the out of sample problem. Experiments on hyperspectral image demonstrated that the proposed method could achieve higher classification rate than the conventional image classification technologies.