In the hyperspectral image classification,the rich data enhance its l and-cover recognition ability.However,the problem of "dimension disaster" is caused by t he large number of sample features and the small number of labeled training sample points.In thi s paper,an unsupervised feature selection based hyperspectral image classification method is proposed,which takes into account both the manifold embedded mapping and sparse representation of the data.The feature selection problem is transformed into an optimization problem,in which the manifold embedding and sparse expression of the data are used as constraint items and add ed to the objective function.Three objective functions are designed,the first objective function describes the local princip le of manifold learning,the second objective function returns the original sample point to the low dimensional embedding space,and the third objective function is regularization of the regre ssion coefficient. Aiming at the non-convex problem of the objective function,an iterative method is proposed to solve this constrained optimization problem,and an algorithm for solving the op timization problem is given.Preferred features are used to participate in subsequent class ification recognition tasks.Experiments on real hyperspectral datasets show that the new method can i mprove the accuracy of classification.