Sparse Manifold Learning For Hyperspectral Imagery

Hyperspectral imagery(HSI)is captured by imaging spectroradiometer record the reflectance values of electromagnetic wave.It contains dozens or even hundreds of contiguous and narrow spectral bands covering the electromagnetic spectrum from visible to near-infrared regions.HSI possesses high spectral resolution,and it can discriminate different land cover types with subtle differences.Now,HSI has been widely used in the fields of environmental monitoring,precision agriculture,target recognition,and land cover classification.For HSI classification,traditional method will result in “the curse of dimensionality” because of huge data,many bands and strong correlation between spectral bands.Therefore,how to reduce the number of bands and preserve some valuable information has been regarded as the research frontier and hotspot in the field of HSI classification.Feature extraction is an effective method to reduce the number of bands in HSI,obtain some desired intrinsic information,and improve the classification result.On the basis of the theories of manifold learning,graph embedding,and sparse representation,this dissertation focuses on the research of sparse manifold learning method for feature extraction of HSI.The main research works are as follows:(1)According to the characteristics and challenges of HSI,the merits of feature extraction were introduced for HSI classification.Then,the feature extraction and classification methods of HSI were introduced,and the development history of feature extraction and classification methods were reviewed.The fundamental theories and corresponding methods of manifold learning,graph embedding,and sparse representation were also introduced,which lays a solid theoretical foundation for this dissertation.Finally,the accuracy evaluation index of HSI classification and some common hyperspectral data sets were summarized.(2)The manifold learning methods were investigated for feature extraction of HSI in this dissertation.With the analysis of manifold learning method,the graph embedding framework was introduced,and the marginal Fisher analysis(MFA)algorithm based on this framework was detailedly analysed.Howerver,MFA cannot effectively represent the intrinsic structure of HSI that possesses many homogenous areas.For the problem,the local geometric structure Fisher analysis(LGSFA)method was proposed in this dissertation.LGSFA utilizes the neighborhood of data and the intraclass reconstruction point of each neighbor point to reveal the intrinsic structure of HSI.It enhances the representation power of feature.The experiments on the Salinas and Indian Pines hypersectral data sets show the effectiveness of the LGSFA method.(3)The feature extraction methods of HSI based on sparse rerepsentation were studied in this dissertation.According to the problem of neighborhood selection for graph construction in the graph embedding framework,the sparsity preserving analysis(SPA)algorithm was proposed.SPA can adaptively reveal the similarity relationships between data points with the nature disctiminating power of sparse representation,and it applies the sparse coefficients to construct a sparse graph.It can extract more effective discriminating feature.For the classification experiments on the PaviaU and Urban hypersectral data sets,SPA obtains better results compared with other corresponding methods.On the basis of the SPA method,the sparse discriminant learning(SDL)method was proposed with the label information of hyperspectral data.SDL uses sparse representation to reveal the similarity relationship between data points.It constructs an intraclass sparse graph and an interclass sparse graph,and it enhances the similarity weights between data points from the same class.The SDL algorithm can improve the separability of data points form different classes.The classification results on the Indian Pines and Urban hypersectral data sets show that SDL can better represent the intrinsic properties of data and improve the classification accuracy.(4)According to manifold learning and sparse representation,this dissertation presented the research on sparse manifold learning.The sparse manifold embedding(SME)method was proposed based on sparse manifold coding that can adaptively choose the data from the same manifold.SME constructs a sparse manifold graph to reveal the sparse manifold structure of data,and it can effectively represent the intrinsic properties of data.With the label information of HSI,this dissertation further proposed the sparse discriminant manifold embedding(SDME)algorithm.SDME enhances the compactness of intracalss data,and it can extract more effective discriminating feature.The experiment results on the Indian Pines and PaviaU hyperspectral data sets show that SME and SDME possesses better classification accuracy compared with other state-of-the-art methods.To utilize the labeled and unlabeled samples of hyperspectral data,the semisupervised sparse manifold discriminative analysis(S3MDA)method was proposed on the basis of sparse manifold coding.According to the sparse coefficients of the labeled and unlabeled samples,S3 MDA constructs a within-calss graph,a between-class graph and an unsupervised graph,respectively.In low-dimensional spase,it compacts the properties of within-class graph,separates the characteristics of between-class graph,and also compacts the similarity of unsupervised graph.S3 MDA can obtain better low-dimensional feature.For the classification experiments on the PaviaU and Salinas hypserspectral data sets,S3 MDA can achieve better accuracy than other feature extraction methods.In summary,this dissertation focus on the research of the feature extraction method of HSI.According to the theories of manifold learning,graph embedding,and sparse representation,this dissertation gradually constructed some HSI feature extraction algorithms,and the effectiveness of each method was demonstrated on some hyperspectral data sets.