Hyperspectral image has been widely used in land-cover classification. Due to huge amount of data and high correlation between bands, band selection technology is main method to reduce computational complexity. According to non-linear relation between band data, Spectral Clustering (SC) is imported to cluster and select band. In this method, neighbor graph and similarity matrix are generated from band image samples, then samples are divided into k clusters by spectral clustering algorithm. At last, k selected representative samples are generated and used in the subsequent classification and recognition task. Experimental results show that new method can reduce computational complexity and improve classification accuracy of land-cover classification.