Dimensionality Reduction Of Hyperspectral Imagery Base On Graph Embedding With Multi-Structure Characterization

Because of its unique three-dimensional data structure,hyperspectral remote sensing image(HSI)contains a large number of feature spectral data,which makes it possible to classify features into pixels,which is an important way to monitor ground targets.However,its extremely high spectral resolution will cause different features to have higher spectral similarity in certain bands,resulting in increased redundant information,and the hyperspectral data have a non-linear distribution phenomenon in the highdimensional space.Classification will consume huge computing resources and easily lead to the phenomenon of "dimensional disaster".Therefore,how to minimize the dimensions of HSI,reduce redundant information in the spectrum,and extract features that are more conducive to classifier classification is one of the hotspots in the research of hyperspectral remote sensing images.In this paper,based on the characteristics of the hyperspectral,such as the small count of labeled pixels and the non-linear distribution,and based on the theoretical basis of manifold learning,graph embedding learning,semi-supervised learning,and sparse representation,a research on hyperspectral image feature extraction algorithms is conducted.The main research contents of this article are:(1)According to the difficulties encountered in hyperspectral image processing,the necessity of dimensionality reduction of hyperspectral images is derived Then it analyzes the current research status of various feature extraction algorithms at home and abroad,and their respective advantages and disadvantages.The feature extraction algorithms based on mathematical statistics,manifold structure,graph framework and sparse expression theory are summarized in detail.Commonly used classification algorithms for hyperspectral images,classification evaluation indicators,and hyperspectral image data sets involved in the article are briefly introduced.(2)A semi-supervised multi graphs embedding(SSMGE)algorithm on hyperspectral feature extraction is proposed.In view of the fact that traditional graph embedding models cannot effectively characterize more complex structures in high-dimensional data using only a single graph structure,and the status that there are few labeled samples in hyperspectral images,a SSMGE algorithm is proposed to maintain the one-to-one relation in graph samples and the multivariate relations in the hypergraph.An intra-class hypergraph,an inter-class hypergraph,an intra-class graph,and an inter-class graph are constructed by labeled pixels.It uses unlabeled samples to construct an unsupervised intrinsic hypergraph and a penalty hypergraph.In the embedding process,the samples in the intra-class graph and the intrinsic hypergraph are closer,the samples in the inter-class graph and the penalty hypergraph are more sparse,and the multiple structural relationships in the data are maintained in the form of multi-graph embedded.On the PaviaU and Urban hyperspectral datasets,comparative experiments are performed with the related theoretical algorithms under the same experimental conditions.The experiments show that compared with a graph or a hypergraph embedding algorithm,SSMGE has a better feature extraction result,which can effectively improve the classification accuracy of the classifier.(3)A local geometric sparse preserving embedding(LGSPE)algorithm on hyperspectral feature extraction is proposed.This method uses a local linear embedding method to reconstruct each sample on the same class to maintain the local linear relationship in the same class data.At the same time,it calculates the local sparse structure in the neighborhood of the sample.Based on this,the local geometric relation and the sparse relation of the HSI are maintained through the graph embedding frame.Finally,it makes the data in the class as close as possible in the low-dimensional embedding space,and uses the obtained low-dimensional identification features to improve the classification effect of the classifier.It can be known by comparing experiments on Indian Pines and PaviaU hyperspectral data show that LGSPE can significantly improve the classification performance of features compared to general local preserving embedding algorithms and manifold learning algorithms.In summary,this paper studies the dimensionality reduction algorithm for the characteristics of hyperspectral remote sensing images distributed in high-dimensional space.The research results show that the algorithms proposed in this article can not only effectively reduce the dimension of hyperspectral data,but also provide new methods and ideas for the dimensionality reduction of high-dimensional data in other fields.It has important theoretical significance and application value.