Hyperspectral Image Spatial-Spectral Feature Extraction And Classification Based On Manifold Learning

Hyperspectral Image(HSI)has abundant spatial-spectral information and is an important basis for accurate classification of ground objects.However,in the process of HSI classification,this kind of high spatial-spectral resolution will lead to large amount of data,much information redundancy and low classification accuracy.Therefore,it is an urgent problem to explore effective dimensionality reduction methods and improve the classification effect when mining the spatial-spectral information of images.In this paper,the nonlinear structure of HSI data is taken as the starting point,the low-dimensional manifold structure of high-dimensional data is revealed based on manifold learning algorithm,the spatial-spectral information of image is mined by different spatial-spectral fusion methods,and the feature extraction and classification algorithm of HSI is deeply studied.The specific work is as follows:1.This paper proposes a spatial-spectral combined PCA-LPP feature extraction algorithm to solve the problem that the local preservation projection algorithm does not exploit the global structural characteristics and underutilizes spatial information.Firstly,based on the principle of spatial consistency,the spatial-spectral combined distance is proposed to measure the similarity of pixels,and the best neighbor point is selected Then the spectral information divergence optimization weight matrix construction method is introduced.Finally,the idea of global structure retention of PCA algorithm is integrated into the optimization goal of LPP algorithm,which makes the low-dimensional space have the characteristics of global linear and local manifold structure retention,effectively extract the identification features,and improve the classification accuracy2.Aiming at the shortcomings of the SSPCA-LPP algorithm that does not use sample label information,a semi-supervised spatial-spectral neighborhood preserving embedding(S2NPE)feature extraction algorithm is proposed from the perspective of maintaining the local reconstruction relationship of the manifold.Firstly,the weighted spatial-spectral combined distance is proposed,the Pearson coefficient is introduced to calculate the spectral distance,and the weighted summation with the spatial-spectral combined distance is used to further mine the spatial-spectral information.Then use a small amount of labeled sample information and local neighbors to set the weights to obtain more effective identification features and improve the classification effect.3.In order to further integrate the spatial-spectral information,a two-stage spatial-spectral classification framework is proposed based on the S2NPE algorithm.First,based on the S2NPE algorithm,one-stage spatial-spectral classification results are obtained.Then,to solve the problem of simple linear iterative clustering algorithm that is susceptible to misinterruption when updating the clustering center,a neighborhood selection SLIC(NS-SLIC)algorithm is proposed,which adds selection constraints during the clustering center update process Reduce the interference of erroneous points.Finally,based on the majority voting method,the clustering result of NS-SLIC and the classification result of S2NPE are subjected to two-stage spatial-spectral fusion to further improve the classification accuracy.On the Indian Pines,Pavia University and Salinas datasets,the algorithm in this paper can effectively mine the spatial-spectral information,achieve dimensionality reduction,and have higher classification accuracy than traditional algorithms.Figure[25]Table[7]Reference[68]...