Limited labeled samples is used adequately as a hard constraint into matrix factorization. Meanwhile, the local graph of data is constructed to exploite the manifold structure and maintain the local invariance. As a result, feature extraction on matrix factorization for hyperspectral data(FEMF)was proposed. Through matrix decompose process, the nearby data in original space is still close after dimensionality reduction, and the congeneric labeled data is projected into the same position. Such low-dimensional representation has more powerful discriminate performance for classification and clustering. This issue is a non-convex programming problem, and an iterative multiplicative updates algorithm is introduced to achieving the local optimization. The efficiency of the proposed method was verified in real hyperspectral data feature extraction.