Research On Hyperspectral Image Band Selection Algorithm Based On Graph Learning

The hyperspectral images（HSIs）have the characteristics of the integration of image and spectrum,high spectral resolution,and a large number of bands,which have been widely used in mineral exploration,ocean detection,vegetation coverage,precision agriculture,and other fields.However,the high redundancy and large data volume of HSIs will increase the computational cost of intelligent image processing algorithms.Therefore,band selection（BS）has become a new focus in HSI processing,which aims to select a representative band subset that illustrates the spectral information of the original HSI data.Based on the theories of graph learning,latent low-rank representation,and band ranking,this paper has successively proposed two HSI BS algorithms.The specific work is as follows:（1）A semi-supervised band selection model by combining the graph learning and band ranking theories（SSBS＿GR）is proposed.Firstly,by calculating the similarity matrix with the unlabeled and labeled samples without supervision,the proposed algorithm can be capable of keeping the similar local association in the original space and low-dimensional space.Then,it applies the limited labeled samples to enhance the discriminativeness of the selected bands.In addition,in order to avoid the high redundancy of the selected bands,L_2;1-norm with row sparsity constraint is utilized to redesign the non-negative indictor matrix,thus making the band subset with representation and interpretability through its ranking.Finally,the experimental results on HSI data show the effectiveness of the proposed algorithm.（2）An unsupervised band selection algorithm based on graph learning and latent low-rank representation（Lat LRR＿GBS）is developed,which integrates the latent low-rank representation,similarity graph learning,and band selection.On the one hand,the local spatial information learning and non-negative constraint conditions are jointly integrated into the latent low-rank representation,which can learn the global and local structure of HIS data in the low-dimensional space by alternately learning with the non-negative indictor matrix.On the other hand,considering that the latent low-rank representation can explore the low-rank property of the data from the perspective of row space and column space,it is applied to the proposed BS model to learn the spatial-spectral information of HSI data.In addition,to enhance the robustness of the proposed algorithm,the error component E with L_2;1-norm constraint is further applied.Finally,the experimental results on HSI data verify the performance of our band selection algorithm.