An unsupervised band selection algorithm for hyperspectral imagery based on maximal information is proposed in this paper. The objective of the method is to preserve the maximal information from original data in the selected bands. The bands with less information are removed one by one from the original data. K-L divergence is used to quantify the information amount and its distribution over all the dataset is considered to judge the specific band which needs to be removed. Compared with traditional methods, the proposed approach has an explicit physical meaning and its computational process is very simple. It is an unsupervised method and can perform automatically.