A semi-supervised feature weighting using graph Laplacian is proposed for hyperspectral band selection. The method first constructs the graph Laplacian modified by the label information. The projection matrix is obtained by solving a generalized eigen-problem. The corresponding matrix coefficients are analyzed using the loading factors to assign weights to the original bands. Experiments with hyperspectral data sets are carried out to make comparison among several band selection algorithms. The results show that the proposed method can achieve the best performance as it makes good use of class information from the labeled samples and local structure clues hidden in numerous unlabeled ones.