Feature Extraction Algorithm For Hyperspectral Images Based On Manifold Learning

Hyperspectral images consists of the electromagnetic wave reflection values of a specific target obtained by the spectral imager.While obtaining the spatial features of the observed target,the spectral features are also obtained,which is the "integration of maps and spectra" of the hyperspectral images.Hyperspectral remote sensing images have high Spectral resolution,multiple bands,and rich ground feature information,but too many bands will lead to too much redundant information and cause "curse of dimensionality".Therefore,when processing hyperspectral images,how to extract effective features while reducing redundant information is a major research hotspot in groun-d object classification of hyperspectral remote sensing images.As one of the effective methods to solve the " dimensionality disaster",manifold learning has played an important role in computer vision,data mining and other fields.Manifold learning can effectively reveal the structure of data.But traditional manifold learning algorithms still have problems such as only utilizing spectral information and ignoring its spatial information.Therefore,this article focuses on how to effectively extract features from hyperspectral images and improve the classification performance of ground objects.The main tasks are as follows:Introduced the development history of hyperspectral remote sensing,analyzed the current research status of feature extraction algorithms at home and abroad,summarized some common classic algorithms,and laid the foundation for future research.In addition,the commonly used classification evaluation indicators and the hyperspectral remote sensing image dataset involved in this article were introduced.The original manifold learning algorithm usually uses the distribution of nearest neighbor points to determine the nearest neighbor structure,resulting in ignoring the global structure only for local structures.Therefore,a hyperspectral dimensionality reduction algorithm based on local divergence preserving projection(LSPP)is proposed for feature extraction in hyperspectral remote sensing images.This algorithm first reconstructs the sample points,and then uses the reconstructed data to construct a divergence matrix to characterize the geometric distribution characteristics of the data.Finally,a target algorithm was proposed by combining the Local Preserved Projection(LPP)algorithm with the divergence matrix.The divergence matrix provides both global and local information,enabling the algorithm to effectively mine nearest neighbor and global structures,enhancing the separability of data in low dimensional spaces.Experiments on the Salinas dataset and PaviaU dataset show that this method can effectively extract the intrinsic manifold structure and improve the classification performance of the algorithm.Due to the existence of nonlinearity,even pixels of the same class may exhibit significant spectral differences,resulting in manifold learning relying solely on spectral information to obtain effective projection information.The addition of spatial information can effectively improve this problem.Therefore,a supervised spatial-spectral joint local preserving projection algorithm(SS-LPP)is proposed and applied to feature extraction of hyperspectral remote sensing images.According to the principle of spatial consistency,the weighted mean filtering algorithm is used to filter the original sample points,and then the reconstructed sample points are used to construct intra class and inter class graphs.By fusing spatial and spectral information,the differences between similar samples are reduced,and the differences between different categories are improved,which can improve the classification accuracy after data dimensionality reduction.Experiments on the Salinas dataset and Pavia U dataset show that this algorithm has better classification performance.