In hyperspectral remote sensing image classification, needs a large number of training samples to trainclassifier, but labeled sanmpes is very difficult ,timeconsuming and expensive.Therefor, we proposed aadaptive method combined representative samples with uncertainty samples to select samples. We use the active learning based on the best vs secondbest(BvSB) for selecting training samples and take advantage of the expectation maximum (Expectation Maximization, EM) cluster to computerepresentativeness.Then uncertainty and representative of the samples combined with aadaptive weight to select most informative unlabeled samples for manual labeling, and join the training set for training classifier. Experiments show that our method is more stable performance and accuracy is also improved.