Hyperspectral image has hundreds of successively narrow bands, which brings serious problems such as large correlation and redundant information. The selection of the optimal bands, which are suited for classification or recognition, has become a difficult work that needs to be overcome. In order to solve the problem of the large correlation among bands, a band selection method based on improved subspace partition through global search on correlation matrix was proposed. Through a global search, the band correlation matrix was divided into a series of subspace, from which the optimal bands were finally selected. The proposed method provides a band selection which has small correlation between each other. The result of an experiment which used Support Vector Machine(SVM) on an AVIRIS image shows that the proposed method is valid.