Feature Extraction Of Hyperspectral Images Based On Collaborative Graph Embedding

Hyperspectral images(HSIs)are acquired by imaging spectrometers from dozens to hundreds of consecutive and narrow bands,which contain spatial information and abundant high-resolution spectral information,enabling the effective identification of subtle differences between different land covers.At present,HSIs have been widely used in various fields,such as fine agriculture,resource inspection,environmental monitoring,biomedicine,etc.Among these fields,the HSI processing technologies are the main research content,where feature extraction is one of the main dimensionality reducation method of HSIs.Its purpose is to find a transformation method to construct the low-dimensional information features,thus preserving the valid information and solving the problem of "Curse of Dimensionality".Inspired by the graph embedding framework,this thesis constructs two collaborative graph embedding-based feature extraction methods for HSIs,and the experimental results conducted on some commonly used HSI data sets verify their effectiveness.The main research work is listed as follows:(1)An HSI feature extraction algorithm based on matrix exponential collaborative graph discriminant embedding(ECGDE)is designed.Firstly,due to the fact that the collaborative graph discriminant analysis algorithm is insufficient for the separability of samples in low dimensional space,based on the collaborative graph,the proposed ECGDE algorithm maximize the inter-class distance and minimize the intra-class distance in the process of projection learning.Furthermore,in order to improve the separability of HSIs in projection space,the matrix exponential is introduced in our ECGDE,maping the scatter matrix that maintains the locality to the exponential space to achieve distance diffusion mapping,and ensuring the positive definiteness of the coefficient matrix in the generalized feature problem.Therefore,the proposed ECGDE algorithm can effectively improve the separability of interclass projection sample,and solve the "Small Sample Size" problem in the feature extraction of HSIs.Experimental results show the superiority and effectiveness of the proposed algorithm.(2)A Tikhonov matrix collaborative representation discriminant neighborhood projection(Tik CRDNP)is designed for the feature extraction of HSIs.The local structure of HSIs is usually ignored in the graph construction process,leading to the inability to obtain effective local information after dimensionality reduction.As such,this thesis first introduces the Tikhonov matrix into the graph construction process to enhance the local structure of samples,enhancing the representational ability of HSI information for the proposed Tik CRDNP method.After that,the projection objective function is further designed with the neighborhood preserving embedding method in our Tik CRDNP so that the samples with the same class can maintain the neighbor relationships of their original high-dimensional space.The experimental results show that the proposed method has better feature extraction ability than other algorithms,thus improving the classification performance of HSIs.