Study Of Hyperspectral Image Dimensionality Reduction And Classification Based On Manifold Metric Learning

Compared with conventional remote sensing technology,hyperspectral remote sensing has a large number of spectral bands and can acquire spectral information of surface objects in tens to hundreds of continuous spectral bands.At the same time,hyperspectral remote sensing can combine the spectral information with the spatial information reflecting the texture,morphology and other characteristics of the surface objects,which has the characteristic of“map-spectrum integration”.The above features make hyperspectral images rich in feature information and have the ability to accurately identify features.However,the large number of bands and narrow band range lead to high similarity between neighboring bands,resulting in a large amount of redundant information in hyperspectral images,which limits the subsequent application of hyperspectral images.Metric learning is a powerful tool to solve this problem,which can reduce the dimensionality of the data while mapping the data to a feature subspace with stronger discriminative power to achieve the purpose of helping hyperspectral remote sensing image classification.In this paper,based on the theory of metric learning,manifold metric learning is proposed to assist hyperspectral image dimensionality reduction,and the specific contents are as follows:(1)This thesis first discusses the importance of dimensionality reduction algorithms in hyperspectral image processing and compares the existing works of scholars at home and abroad,then introduces the basic principles and representative algorithms of manifold learning and metric learning,as well as the hyperspectral datasets used for experiments and the evaluation metrics.(2)To address the problem that the existing metric learning ignores the mining of data neighborhood features,which leads to the inability to distinguish highly similar heterogeneous samples when facing heterogeneously distributed data,this paper proposes the Clustered Multiple Manifold Metric Learning(CM~3L)algorithm.CM~3L algorithm combined with multi-task learning,by dividing the original training data into multiple independent clusters and treating each cluster as a separate training task.Then the sample points in the clusters and their neighbors are considered as a whole and constructed as manifold.Finally,the construction of similarity constraint and dissimilarity constraint is completed according to the manifold distance,and the features with high discriminative power are mined by using the local neighborhood feature information of hyperspectral images.On three real hyperspectral datasets,KSC,Houston and University of Pavia,CM~3L classification accuracy reaches 87.12%,94.17%and 91.11%,respectively,which outperforms the comparative methods such as manifold learning and metric learning under the same conditions.(3)To address the problem that the manifold distance ignores sample points in unlabeled data when constructing manifold structures,resulting in incomplete manifold structures and affecting the accuracy of manifold distance calculation,this paper proposes the Spatial-Spectral Manifold Distance Metric Learning(SSMD)algorithm.Based on the distance of the manifold in CM~3L,SSMD selects the appropriate nearest neighbors in the labeled and unlabeled data to participate in the construction of the manifold based on the spatial-spectral information,which ensures the completeness of the constructed manifold.Then the distances between the manifolds are calculated to replace the traditional Mahalanobis distance,and finally the constructed SSMD are extended to other metric learning to improve the feature extraction capability of metric learning in hyperspectral dimensionality reduction and classification.The SSMD was used on three real hyperspectral datasets,KSC,Houston and University of Pavia,and achieved better classification accuracy compared to the original algorithm,with the maximum improvement of 4.23%,3.58%and 4.96%,respectively.