Because the existing unsupervised band selection methods in hyperspectral image classification could not calculate the similarity between bands and the high-dimensional characteristics in the selection process， a greedy unsupervised hyperspectral band selection method based on variable precision rough set was proposed. Firstly， a new dependency measure was defined by using the variable precision rough set， which made it insensitive to the misclassification parameters of the variable precision rough set， so as to make full use of the similarity between wavebands. Secondly， a new criterion was proposed to find out the bands with higher and lower similarity values in the unselected and selected segment subsets. Then， the first order incremental search method was used to select the required information band one by one， so as to avoid the generation of a large amount of information and reduce the computational complexity. Finally， three hyperspectral datasets were used to compare the proposed band selection technique with the five latest techniques. The results show that the proposed method has good classification accuracy for all datasets， and the average classification accuracy is only 1.9%，3.1% and 4.1% lower than the average classification accuracy of all pixels under the condition of 50% marked pixels， which proves that the proposed method can guarantee good classification performance and generalization ability of data set， and has robustness to parameters.