Aiming at the problem that the number of hyperspectral image bands is huge and it is easy to cause "dimension disaster," an improved K-means hyperspectral classification algorithm based on variance coefficient weighting is proposed. The algorithm is based on the traditional K-means algorithm, it portray the importance of bands by introducing variance coefficients and express intra-class correlation using correlation coefficients. Finally, it incorporates inter-class information to achieve global optimal clustering. The algorithm in this paper compensates for the shortcomings of the K-means algorithm in viewing each feature equally, while taking into account intra-class and inter-class information, which enhances the accuracy of hyperspectral image classification to a certain extent.