Dimensionality Reduction Method For Hyperspectral Imagery Based On Spatial-Spectral Manifold Learning

Remote sensing acquires the electromagnetic wave characteristics of target objects by modern optical and electronic detection instruments from a long distance,so as to analyse the shape,position,state and other changes of the target object through information transmission,correction and interpretation.Hyperspectral remote sensing image(HSI)is composed of the reflection value of electromagnetic wave detected by hyperspectral sensors,it contains abundant spatial information,radiation information and spectral information,which can accurately discriminate the subtle differences between different land-cover types.HSI has been widely used in the fields of precision agriculture,mineral exploration,target recognition,environmental monitoring,and urban planning.However,the large volume and high redundancy of HSIs has brought huge challenge to land cover classification.Therefore,how to reduce the dimensionality of hyperspectral data while preserving the useful information is one of the most important issues in the field of hyperspectral remote sensing.Manifold learning can reveal the intrinsic manifold structure in high-dimensional data,and it has achieved good performance in biometric recognition,computer vision,data mining and other fields.However,when applied to hyperspectral remote sensing images,manifold learning methods have problems such as inability to reveal the multimanifold structure and insufficient use of spatial-spectral combined information.Therefore,this dissertation conducts the research of dimension reduction(DR)in spatialspectral manifold learning for HSI classification.It can explore the intrinsic multimanifold structure of HSI and improve the performance of land cover classification,which has significant scientific and practical application values for many fields.The main contributions of this paper are as follows:(1)Considering that the single adjacency graph cannot effectively represent the neighbor relationship between data points,an unsupervised local neighborhood structure preserving embedding(LNSPE)algorithm was proposed in this dissertation.At first,LNSPE reconstructs the samples with their spectral neighbors.Then,it constructs a novel adjacency graph structure by original data and reconstructed data,which can effectively characterize the local neighbor relationship between data points.On this basis,it adopts the total scatter matrix to characterize the distribution characteristics of the HSI data,and combines the dual graph structure and the scatter information to enhance the separability of data points in low-dimensional space.Experimental results on the public HSI data sets of Pavia U and Botswana show that the LNSPE algorithm can effectively improve the classification accuracy.(2)To effectively utilize the spatial information of HSI data,an unsupervised spatial-spectral manifold reconstruction preserved embedding(SSMRPE)algorithm was proposed based on the spatial distribution consistency of HSI.At first,a spatial-spectral combined distance(SSCD)was designed for selecting the spatial-spectral neighbors of samples,which can construct an effective spatial-spectral adjacency graph to reveal the intrinsic manifold structure of HSI.Then,it defined a new error loss function based on the spatial coordinate relationship of data points,which can adjust the reconstruction coefficient adaptively to extract low-dimensional embedding features.The experimental results on Pavia U and Salinas hyperspectral datasets demonstrate the effectiveness of the SSMRPE algorithm.(3)To reveal the intrinsic multi-manifold structure in HSI,a multi-manifold locality graph preserving analysis(MLGPA)algorithm was proposed.MLGPA takes each class of hyperspectral data as a submanifold,and it establishes the relateionship of data samples on each submanifold by constructing intramanifold graph and intermanifold graph,and each submanifold can obtain a projection matrix from high-dimensional space to lowdimensional space.After that,MLGPA effectively fuses the embedded features on each submanifold to enhance the representation ability of low-dimensional features and improve the classification performance.To further utilize the spatial-spectral combined information of HSI,a spatial-spectral multiple manifold discriminant analysis(SSMMDA)algorithm was proposed.For each submanifold,SSMMDA constructs a spectral-domain intramanifold graph to reveal the local neighborhood structure of HSI,and then it constructs a spatial-domain intramanifold graph and a spatial-domain intermanifold graph to characterize the within-manifold aggregation and betweenmanifold separation,respectively.Finally,it designs a spatial-spectral combined objective function to obtain the optimal projection direction of each submanifold,and extract the effective spatial-spectral discriminant features.The experimental results on Pavia U and Washington DC Mall hyperspectral data sets demonstrate the effectiveness of the proposed MLGPA and SSMMDA methods.(4)To further verify the practicability and effectiveness of the proposed algorithms,this dissertation carried out dimensionality reduction experiments on two domestic hyperspectral remote sensing images,and NN and SVM classifiers are used to verify the classification performance of embedding features.Furthermore,the characteristics and applicability of each algorithm are analyzed in this dissertation.