Research On Feature Extraction Of Hyperspectral Imagery Based On Graph Embedding

Hyperspectral imagery(HSI)is usually composed of tens to hundreds of spectral bands,and reflectance values of a pixel at every bands constitute its spectral curve.Generally,pixels of the same type have similar spectral curves,while spectral curves of different types of pixels have differences.From this,different types of pixels can be distinguished,and the intrinsic properties of ground targets can be revealed.Therefore,HSI can not only be used for land-cover classification,but also help to discover mineral content,soil moisture,vegetation health,the physical composition of buildings,and other invisible details.However,while providing rich information,high dimensional data also has Hughes phenomenon.How to extract effective features from high dimensional data for HSI classification has attracted the attention of many researchers.Based on the theory of graph embedding,manifold learning and deep learning,the spectral feature extraction of HSI is studied in this paper.The main work can be summarized as follows:(1)The development of hyperspectral imaging sensor for earth observation and the progress of intelligent remote sensing are introduced.Then,the research status of HSI feature extraction is analyzed,and several classic feature extraction methods are introduced as basis for subsequent research.In addition,three kinds of hyperspectral scenes and related evaluation indexes are also mentioned in this paper.(2)A local reconstruction Fisher analysis(LRFA)method is proposed for feature extraction of HSI.Local geometric structure Fisher analysis(LGSFA)uses original pixels and reconstruction pixels to construct adjacency graphs together,which can not effectively preserve the global structure of nonlinear manifolds in the low dimensional space.To address this issue,LRFA first reconstructs each original pixel from its intraclass neighbors.And then intrinsic graphs and penalty graphs are constructed based on reconstructed pixels.In the low dimensional feature space,it not only preserves the integrity of manifold,but also improves the separability of features.Good discriminative features can improve the classification performance of HSI.Experiments on two hyperspectral remote sensing scenes of Pavia University and Urban show that,compared with LGSFA method,the proposed LRFA significantly improves classification accuracy and saves running time.(3)In the face of HSI with complex nonlinear structures,a deep feature Fisher analysis(DFFA)method is proposed to help to extract abstract features and improve the separability of deep features.DFFA first obtains deep features of HSI from an unsupervised autoencoder.Then intrinsic graphs and penalty graphs are constructed to reveal intrinsic geometric structure of pixels.Finally,the deep features are mapped into a low dimensional space,in which distance between pixels from the same class becomes smaller,and distance between pixels of different classes increases.The feasibility and effectivity are verified by experiments on three hyperspectral remote sensing scenes of Pavia University,Indian Pines and Urban.Compared with manifold learning and deep learning,DFFA improves feature extraction ability and classification performance.