Research On Feature Extraction And Classification Recognition Algorithm For Hyperspectral

Hyperspectral imagery is a high-dimensional data that contains abundant spatial,spectral,and radiation information.It is widely used in geological mapping and exploration,atmospheric or vegetation ecological monitoring,product quality inspection,precision agriculture,urban remote sensing,and military battlefield reconnaissance.The hyperspectral remote sensing images have the characteristics of high redundancy,and strong correlation and large data volume,and these features bring many challenges to the classification and identification of hyperspectral images.The feature extraction and classification methods of hyperspectral images are systematically studied in this paper.The texture features of hyperspectral image bands,and the Euclidean distance between bands,and the dimension reduction methods of hyperspectral images based on clustering are discussed are investigated in detail.In addition,we have studied in depth the intra-class band index and support vector machine classification model,and focus on hyperspectral image classification method based on the clustering and intra-class band index and its simulation experiments.The main work and contributions of the thesis are as follows.(1)The characteristics of hyperspectral image data are studied,and the related theoretical basis of hyperspectral image feature extraction algorithm and classification recognition method in recent years are learned.The principle of maximum variance principal component analysis,and adaptive band selection method and automatic subspace partition method are emphatically studied.The above methods are verified by simulation experiments.(2)A hyperspectral band selection method based on multi-features and affinity propagation clustering algorithm is proposed.This method can overcome the problem of constructing similar matrix with single information and the difficulty of selecting initial values in some clustering methods,and the problem that given initial cluster centers have a great impact on clustering results.In the algorithm,the texture features of the band image are firstly obtained.Because Euclidean distance can describe the spatial distance relationship between bands and it is insensitive to the difference in spectral amplitude,we combine the band texture features with the Euclidean distance to construct a similarity matrix.Then the affinity propagation clustering algorithm is used to cluster the hyperspectral band image,and the cluster center of each class is selected as the band representation.Thus,the band reduction dimension of the hyperspectral image is achieved.The experimental result shows that the method is feasible.(3)A hyperspectral image classification method based on clustering and intra-class band index is proposed.Firstly,after analyzing the characteristics of bands in clusters and combining the strong correlation of the adjacent bands in hyperspectral image,a method to remove noise bands is proposed.Then the advantages and disadvantages of the optimal index factor and the adaptive band index are studied and analyzed.A new band index calculation method is proposed,named as intra-class band index,which fully considers the information content of the band and the correlation between the bands.The support vector machine model and the problems of Lagrangian solving convex optimization are analyzed and derived.The support vector machine based on radial basis function kernel is used for classification.For the parameter selection problem,the grid parameter optimization method is adopted,and the V-fold cross-validation method is used to measure the model generalization ability and avoid the over-fit problem during parameter optimization.(4)The simulation experiments of two hyperspectral images are performed.Then the results are analyzed from the aspects of the overall classification accuracy,Kappa coefficient,confusion matrix and user accuracy.In addition,in order to verify the performance of the method proposed in the paper,a number of repeated simulation experiments are performed and various comparison method box plots are drawn.It can be seen from the results that the proposed method has good classification effect,stable performance and overall superiority.