Research On Band Selection Method For Hyperspectral Remote Sensing Image Based On Clustering

Hyperspectral Remote Sensing Image(HSI)can provide a more accurate basis for feature identification and classification because of its rich spatial and spectral information.As the spectral resolution increases,a large amount of high-dimensional data is generated during the imaging process,which inevitably causes dimensional disasters and is not conducive to image storage and processing in later stages.Hyperspectral band selection,as an effective way to reduce data redundancy,aims to select a small number of representative feature bands from the original HSI as its optimal subset of bands.However,the spatial resolution,spectral characteristics,feature size,and target scene of HSI are complex and variable.These challenges make band selection one of the most challenging problems in HSI image processing.There are several challenges in band selection:(1)Existing algorithms usually stretch each band directly into a single feature vector during pre-processing hyperspectral band data,ignoring the impact of feature variability in the HSI on the band selection.(2)The existing algorithms propose to use the original pixel information in HSI to represent the bands,however,the pixel features are susceptible to noise interference and there is a lot of redundant information,which significantly limits the ability to represent the band features.(3)In the spatial distribution of features in HSI,the existing algorithms only consider the correlation of bands within a local area,but ignore the correlation of bands under non-adjacent areas.Therefore,in order to solve the above problems,the following band selection algorithms are proposed in this thesis,respectively.(1)This thesis proposes a band clustering and selection algorithm for latent feature cross-fusion to address the low feature variability utilization of existing methods.The algorithm first uses Principal Components Analysis(PCA)and Entropy Rate Superpixel Segmentation(ERS)to segment the first principal component of the HSI into several regions to fully retain the null spectral information it contains,and then learns from each The band is then characterized by learning the corresponding low-dimensional potential features from each segmented region.A unified feature representation of the HSI is generated using a feature cross-fusion strategy,and k-means clustering is performed based on this feature representation to divide all bands into clusters.Finally,the band closest to the cluster central is selected as the feature band from each cluster,thus achieving optimal band subset selection.The proposed algorithm achieves superior performance compared to some current state-of-the-art band selection methods.(2)This thesis proposes a multi-region representation-enhanced band clustering and selection algorithm based on the above-mentioned feature cross-fusion algorithm to address the insufficient representation power of band features in existing methods.Unlike the above algorithms,this algorithm uses a hierarchical strategy to learn the corresponding low-dimensional potential features from each segmented region,which in turn enhances the band feature representation.Finally,the band is subsetted by constructing a unified representation of the HSI and clustering it,and then the band with the highest information entropy from each subset is selected as the feature band,thus completing the band selection.The experimental results on four public datasets show that the proposed algorithm can achieve better performance in band selection.(3)To address the poor modelling capability of the null spectral structure of existing methods,this thesis proposes a band selection algorithm based on graphconstrained subspace clustering of the null spectrum.The algorithm first uses PCA and ERS to preserve the spatially local geometry of the HSI and introduces a self-expressive subspace clustering model to capture the band correlations under this region.In addition,considering that the bands under non-adjacent regions are also somewhat correlated,the proposed method constructs a regional similarity map under the global structure to constrain the model a priori.By coupling the local and global structure information into a unified framework,the spatial-spectral structure modelling capability of the proposed model is further enhanced.Finally,the generated band correlation coefficient matrices are spectrally clustered to complete the partitioning of all band subsets,and then the band with the highest information entropy from each subset is selected as the characteristic band to achieve band selection.The experimental results on three common data sets show that the proposed algorithm can effectively solve the band redundancy problem.