Abstract: Using the least manually labeled samples to train a good classifier is a key problem in image classification．Aiming at selecting these samples for labeling,this paper proposed a criterion combining two different measures of samples:uncertainty of classification and representativeness．The best vs．second-best (BvSB) method is used to get the measure of uncertainty．The dataset is first hierarchically clustered and then the measure of representativeness of each unlabeled sample is defined based on the structural information of clusters and the distribution information of those labeled samples．The proposed method was compared with the random-selection method and BvSB method on an optical image dataset and a fully-polarimetric synthetic aperture radar (SAR) image dataset.The results show that it has stably better performance．

Key words: Active learning,Hierarchical clustering,Image classification