Band selection is an effective method for dimensionality reduction. However, the information from the small size of labeled samples usually misleads the supervised band selection. A semi-supervised band selection method based on graph Laplacian and self-training idea is proposed. The method first puts forward the graph-based semi-supervised criterion for feature ranking to generate the initial band subset. The graph Laplacian used in the criterion is refined with aid of the label information. Then, the supervised classifications are carried out based on the band subset and some unlabeled samples with higher confidence values are added into the labeled sample set. Afterwards the band subset is updated according to the feature ranking based on the newly generated labeled and unlabeled data, and is used for classification. The process repeats to obtain the final subset. The experiments on hyperspectral data sets are carried out compare several unsupervised, supervised and semi-supervised band selection methods. Results show that the proposed method can produce the band subset with better performance.