In order to make full use of the abundant spectral and spatial information of hyperspectral images, a novel feature selection algorithm based on the structure preserving combination with the multi-scale spatial filtering and the hierarchical network is proposed. The feature subset that best preserving the global similarity and the local manifold structure is selected via l2,1 norm mathematical model. The bilateral filtering with multi-scale window and adaptive parameter setting is used for incorporating spatial information into spectral data automatically, enhancing the similarity within class and dissimilarity between different classes. The hierarchical network is introduced to achieve further integration of spatial and spectral information that benefit the classification. The influence of the hierarchical network depth and spatial filtering scale number is analyzed. The experiments validate the effectiveness of the algorithm. The overall classification accuracies reaches to 90.98% and 94.20% on Indian Pines and PaviaU data sets respectively, which significantly improve the classification of land cover compared with conventional methods.