Research On Hyperspectral Image Classification Based On Canonical Correlation Analysis

Hyperspectral image is a three-dimensional cubic image that combines spatial information and spectral information.With the development of hyperspectral remote sensing technology,hyperspectral images are widely used in the fields of land cover classification,agricultural testing,resource investigation,environmental assessment,and urban research.Due to the characteristics of high spectral resolution,lots of spectral information,and rich spectral information,it can accurately detect and distinguish small differences between land cover categories.The classification of hyperspectral images has gradually become an important research field.In the research of hyperspectral image classification,the algorithm of combining spectral and spatial information for classification has been relatively mature.But there are still three problems: 1)the high dimensionality and the large amount of computation,2)general dimension reduction methods are easy to lose information,3)the potential correlation between spectral and spatial information is ignored in the process of dimensionality reduction.To solve these problems,this paper proposed three composite kernel frameworks based on canonical correlation analysis for hyperspectral image classification.1.In order to fully exploit the potential relationship between spectral and spatial information,this paper proposes a composite kernel framework based on canonical correlation analysis for hyperspectral image classification.The typical correlation analysis is introduced to process the spectral information and spatial information to obtain two sets of information with less redundancy and dimensionality reduction,then use the kernel method for information fusion,and finally input the classifier.Through experiments on two sets of real hyperspectral data sets,it is confirmed that the algorithm proposed in this paper can effectively improve the classification accuracy compared with the general composite kernel framework,and it is more robust in the case of small training sample sets.2.Considering the dimensional characteristics of hyperspectral images,this paper further proposes a hyperspectral image classification algorithm based on a two-dimensional canonical correlation analysis composite kernel framework.At first,the proposed algorithm extracts the spectral information and spatial information,then processes each pixel information of the hyperspectral image to convert it into two-dimensional data,and then introduces two-dimensional typical correlation analysis to re-express the spectral information and spatial information.Finally use the compound kernel for information fusion and input the classifier for classification.Through experiments on two sets of real hyperspectral data sets,the effectiveness of the proposed method is confirmed.3.Finally,in order to fuse multiple features including spectrum and multiple spatial information,based on the multiple kernel learning method,this paper proposes a hyperspectral image classification algorithm based on multiset canonical correlation analysis.While using the spectral information,the proposed method includes spatial information of various characteristics,and then introduce multiset canonical correlation analysis for processing.Experimental results show that the proposed algorithm can fully exploit the potential relationship between spectral information and spatial information,reduce information redundancy,achieve data dimensionality reduction,reduce training time,and improve the accuracy and robustness of existing classification algorithms.