Research On Semi-supervised Feature Extraction Of Hyperspectral Image Based On Spatial And Global Structure

The rise of remote sensing technology can comprehensively,stereoscopically and efficiently reflect the distribution of resources in different fields.It has become a new method and new means for human beings to understand the world from a multidimensional and macroscopic perspective.At the same time,the development of hyperspectral remote sensing benefits from the development and maturity of remote sensing imaging technology.Hyperspectral images have a large amount of information and have different penetrations of different bands to various types of features.So as to it can excavate the detailed information of the features.Of course hyperspectral images have the speed of updating,dynamically monitoring the environment and studying the discipline of changes in features.At the same time,in the process of obtaining information,it has the characteristics of limited conditions,wide application and high efficiency.Therefore,many harsh environments that humans cannot conduct on-the-spot investigations can obtain timely and valuable remote sensing image data through remote sensing technology.Hyperspectral remote sensing images have become the focus of human research.However,there are still difficulties in accurately obtaining important features from hyperspectral images and achieving efficient classification.Due to the high frequency spectrum of the hyperspectral image and the phenomenon of information redundancy and dimensionality disaster,the classification accuracy is often low.Therefore,the feature extraction pre-processing operation before hyperspectral image classification processing is crucial.In this paper,hyperspectral image data with large amount of information,strong inter-band similarity and image spatial correlation are taken as research objects.The semisupervised feature extraction for hyperspectral image data is studied with the classification accuracy of image data as the target.The method is mainly carried out from the following three aspects:(1)The imaging principle and band characteristics of hyperspectral imagery are summarized.At the same time,the feasibility and effectiveness of the existing feature extraction method are analyzed due to the multi-dimensionality of hyperspectral image and the high redundancy of information.The acquisition of hyperspectral image class information is difficult and costly.However,due to the further development of semisupervised learning,the process of semi-supervised extracting the features of hyperspectral image is easier and the extraction cost is greatly reduced.The rate and classification effect of extraction features are improved.(2)A semi-supervised LPP feature extraction method of hyperspectral image based on spatial correlation is proposed.The features of the hyperspectral image are extracted according to the local preserving characteristics of the LPP algorithm,but the LPP algorithm only considers the spectral similarity of the image.Hyperspectral images have both spectral information and spatial information,and hyperspectral images have the phenomenon of different spectra with same objects and same spectra with different objects.Therefore,based on the semi-supervised idea,a small amount of class information,spectral information and spatial information of image data are used to construct the pixel weight relationship considering different spectra with same objects and same spectra with different object,and then the weight relationship is embedded in the local preserving objective function.The objective function is solved by the Lagrangian multiplier method,and the feature vector of the solution is used to transform the hyperspectral image data.The experimental results show that the proposed algorithm greatly extracts the structural information of hyperspectral image considering the image spatial information.Compared with the supervised feature extraction,unsupervised feature extraction and semi-supervised feature extraction,the classification accuracy of the features extracted by the algorithm is improved.(3)A semi-supervised feature extraction method of hyperspectral image based on global and local structure is proposed.The global structure and local structure are equally important for the analysis of hyperspectral images.The global structure reflects the overall information of the image,and the local structure reflects the local information of the image.Based on the semi-supervised LPP algorithm considering spatial correlation,the proposed algorithm considers the global structure of hyperspectral images.It is divided into three steps.Firstly,the divergence matrix of the LDA algorithm is used to mine the global intra-class discrimination and the global inter-class discrimination information of the labeled samples,and the semi-supervised PCA algorithm is used to preserve the global structure of the training sample data.Secondly,due to the problem of difficult selection of neighborhood parameters in semi-supervised LPP algorithm,this paper adaptively selects the local manifold neighbors between training samples by sparse representation optimization model,and combines the local intra-class discriminant weights with the local inter-class discriminant weights to save the local structure of training data.Finally,the two parts are embedded in the objective function of the global discriminant and local sparse preserving semi-supervised feature extraction algorithm,so that the features are extracted to obtain more effective and representative features.The experimental results show that the proposed algorithm extracts the feature information of hyperspectral image better while considering global information and local information.