Dimensionality Reduction For Hyperspectral Remote Sensing Image Based On Sparse Multi-manifold Learning

The hyperspectral remote sensing images contain abundant spatial information and spectral information.Due to its characteristics of large numbers of data,many bands,strong correlation between bands and high redundancy,to deal with data with traditional methods is easy to fall into “curse of dimensionality”.Therefore,how to reduce the dimension of the hyperspectral remote sensing images without the loss of useful information,and improve the classification accuracy,has become one of the key problems of the hyperspectral remote sensing image classification.According to the characteristics of hyperspectral remote sensing data,combined with sparse representation,manifold learning and graph embedding,this paper is focused on the research of dimensionality reduction algorithms for the hyperspectral remote sensing images based on sparse manifold learning.The main research works are as follow:(1)To analysis the characteristics and application fields of hyperspectral remote sensing images,and introduce the dimensionality reduction algorithms and the current research status briefly.The principle theories and corresponding algorithms of manifold learning and sparse representation are introduced,which provides a theoretical foundation for this paper.Finally,the accuracy evaluation indexes of hyperspectral remote sensing image classification and several common hyperspectral remote sensing data sets are summarized.(2)Supervised sparse manifold embedding algorithm for dimensionality reduction of the hyperspectral remote sensing images is proposed in this paper.Sparse manifold clustering and embedding(SMCE)can't deal with the new samples directly,and the unsupervised nature restricts its discriminating capability without using the labeled information.For the shortages of SMCE,the supervised sparse manifold embedding(SSME)algorithm is proposed.SSME algorithm obtains the sparse coefficients in affine subspace by sparse manifold coding(SMC).It constructs a similarity graph by sparse coefficients based on graph embedding,and exploits the label information to increase the similarity between homogeneous data and improve the compactness between homogeneous data.The experiment results on Pavia U and Urban hyperspectral remote sensing data sets show that SSME can represent the intrinsic properties and improve classification accuracy.(3)Semi-supervised sparse multi-manifold embedding algorithm for dimensionality reduction of the hyperspectral remote sensing image is proposed in this paper.According to the sparse representation and manifold learning,combined with graph embedding and semi-supervised learning,the semi-supervised sparse multi-manifold embedding algorithm is proposed.It utilizes the useful information from both labeled and unlabeled samples to find the sparse coefficients based on SMC,and adaptively obtains the data points from the same manifold,which can construct a sparse graph to represent the multi-manifold structure of hyperspectral data.Then it tries to exploit label information to increasing the compactness of the data that lie in the same manifold to obtain the low dimensional features on each manifold.Then the effectiveness of S~3MME is verified by experiments on hyperspectral remote sensing data sets Pavia U and Salinas.In summary,according to the characteristics of hyperspectral remote sensing data,this paper proposes two new dimensionality reduction algorithms for hyperspectral remote sensing images,and verify their effectiveness by experimenting on different hyperspectral remote sensing data sets.