Research On Unsupervised Hyperspectral Band Selection Based On Adaptive Multiple Graph Learning

The selection of hyperspectral bands is an important research direction in the field of hyperspectral image processing.It can effectively reduce the data volume of hyperspectral im-ages while preserving their main information,which is helpful for subsequent application tasks of hyperspectral images.In general,hyperspectral images contain rich spectral and spatial in-formation.However,in most previous research work,only the spectral information was used to select bands,such as sorting bands based on the amount of information contained within them and using clustering algorithms to cluster bands.And few studies have considered the spatial information in the process of band selection.Therefore,the information in hyperspectral im-ages has not been fully utilized.Graph is a commonly used data structure for describing spatial information,but in traditional graph learning methods,the construction of the graph is single,and the constructed graph is not embedded in the learning process.For this reason,this paper has carried out in-depth research on unsupervised hyperspectral band selection methods based on adaptive multiple graph learning.Based on the theory of low dimensional subspace projec-tion and spectral clustering,two effective and efficient hyperspectral band selection methods have been designed and proposed by combining adaptive multiple graph learning with sparse principal component analysis,local manifold preserving,and superpixel segmentation tech-nology.These two methods have achieved excellent performance in terms of classification performance of selected bands and algorithm efficiency.The main research content of this article can be summarized as follows:（1）According to the idea of low dimensional subspace projection,this paper proposes a band selection method based on sparse principal component analysis and adaptive multiple graph learning.This method assumes the existence of a low dimensional orthogonal subspace,and the original high dimensional data can be mapped to that subspace through a projection matrix,which reveals the importance of each original band in constructing low features.This method learns the optimal projection matrix through a sparse principal component analysis term with_2,1sparse regularization constraint.In order to maintain the local neighborhood structure of low dimensional data,this method proposes a local manifold preserving constraint term to achieve this goal,and proposes an adaptive multiple graph learning strategy that con-tinuously optimizes the similarity matrix during band selection to enhance the effectiveness of the local manifold preserving constraint.Finally,the performance of the selected bands and the algorithm efficiency of this method are verified on two popular hyperspectral image datasets.（2）According to the idea of spectral clustering,this paper proposes a band selection method based on superpixel segmentation and multiple graph fusion.Considering that differ-ent objects in hyperspectral images have different reflection characteristics,this method pro-poses to use superpixel segmentation algorithm to explore the spatial information of different regions.First,the hyperspectral image is segmented into multiple superpixel regions using the superpixel segmentation algorithm,and then the similarity graph between bands is constructed on each superpixel region.Then,these similarity graphs are fused into a more complete and accurate similarity graph between bands through the adaptive multiple graph learning strategy.Finally,the spectral clustering algorithm is used to select the best band subset on the similarity graph.In addition,the performance of the selected bands and the algorithm efficiency of this method are verified on two popular hyperspectral image datasets.