Manifold Dimensionality Reduction For Hyperspectral Image Based On Spatial-spectral Covariance Feature

Hyperspectral images contain hundreds of spectral bands,which provides rich spectral information for high-precision identification of ground objects,but also brings about the curse of dimensionality.How to reduce the dimensions of hyperspectral images and retain their useful information as much as possible has become a research hotspot in recent years.Compared with the traditional dimensionality reduction method based on statistical theory,manifold learning can well exploit the potential nonlinear structure of high-dimensional data and is widely used in the dimensionality reduction of hyperspectral images.However,the common manifold learning methods generally have the following two problems: the pixels of the hyperspectral image are regarded as isolated points in the high-dimensional space,and the spatial information of the image is ignored;affected by the phenomenon of“different objects with similar spectrums”,it is easy to get undesired local characteristics of the manifold when the spectral features are used to find the nearest neighbors of data points under the Euclidean distance metric.To solve the problem mentioned above,we introduce the spatial-spectral covariance feature to characterize the pixel of hyperspectral images,based on which,three dimensionality reduction methods are proposed in this thesis.The main research works of this thesis are concluded as follows:1.Through theoretical analysis and experiments,from two aspects of feature expression and similarity measurement,this thesis clarifies that spatial-spectral covariance features can be used to learn more accurate relationships among samples than traditional spectral features,so as to recover the data manifold which is more similar to the real one.Spatial-spectral covariance features can make full use of the spatial and spectral information of hyperspectral images,effectively alleviating the influence of the phenomenon of“different objects with similar spectrums”;and the corresponding Log-Euclidean distance metric can be used to learn more accurate relationship between samples by calculating the geodesic distance between the corresponding two points on the manifold.2.To solve the problem that the traditional LLE algorithm can not find reliable samples’ nearest neighbors,local Riemannian embedding for hyperspectral image based on spatialspectral group covariance feature(GLRE)is proposed in this thesis.GLRE adopts a hierarchical strategy,and divides the hyperspectral image into several homogeneous parts based on the disjoint information among bands,with a clustering method,reducing the complexity of the corresponding manifold.In each part,the spatial-spectral covariance features and the local tangent planes of the corresponding Riemannian manifold are used to mine the local characteristics of the potential manifold and extract the discriminant information.The classification results of experiments on real datasets prove the effectiveness of GLRE.3.To solve the problem that the traditional LPP algorithm applies a unified projection on the entire hyperspectral image to reduce the dimensions,ignoring the local homogeneity of the image,superpixelwise locality preserving projection for hyperspectral image classification based on spatial-spectral covariance feature(SuperLPP-SSCF)is proposed in this thesis.SuperLPP-SSCF algorithm divides the hyperspectral image into many sub-regions according to the local homogeneities of hyperspectral image,and then utilizes LPP based on spatial-spectral covariance feature to reduce dimensions of data in each sub-region,which enhances the discriminability of the obtained low-dimensional features.To solve the problem that the number of superpixels in SuperLPP-SSCF algorithm is difficult to determine,multiscale superpixelwise locality preserving projection for hyperspectral image classification(MSuperLPP)is developed.MSuperLPP algorithm integrates the classification results from multiple scales of SuperLPP-SSCF using the decision fusion strategy of majority voting to improve the classification accuracy.The experimental results on Indian Pines and Zaoyuan datasets show that the two methods can promote the hyperspectral image classification.