The traditional Feature Selection methods of hyperspectral data include supervised and unsupervised modes, is not efficient for the real processing system in which abundant unlabeled and few labeled data coexist. Additionally, most of existing feature selection methods ignore that real data has a manifold structure which embedded into the high dimensional space. In order to solve these problems, a Manifold based Semisupervised Feature Selection (MSFS) algorithm was proposed. Considering the prior information of labeled data with the local and nonlocal invariance of the whole data, the discriminate structure is optimized through simultaneously maximizing betweenclass and minimizing withinclass variances. Meanwhile, the manifold structure is exploited from constructing local and nonlocal graphs for the whole data. Then, the efficient features is selected by defining an appropriate evaluation criterion. Finally, through performing the classification experiment on the selected features of real hyperspectral data, it demonstated that our method is able to retain the main structure of data after dimensionality reduction well.