Research On The Semi-supervised Slassification Of Hyperspectral Images With Clustering Algorithm

With the development of imaging spectroscopy technology,hyperspectral remote sensing technology has been developed and applied in many fields.Hyperspectral remote sensing technology combines the traditional imagery space dimension and spectral dimension together,and has rich information of terrestrial spectral information.Therefore,the classification and clustering of hyperspectral imagery data have become the two main directions of hyperspectral data processing.Today's hyperspectral imagery classification and clustering techniques tend to be isolated,and techniques or algorithms that combine the advantages of both are rare.For this reason,a new semi-supervised classification algorithm of hyperspectral imagery based on clustering features is proposed.The work done in this paper is described as follows:Firstly,the important role of hyperspectral imagery in the field of remote sensing technology is expounded.The imaging theory of hyperspectral imagery and its data characteristics are analyzed.The history and current situation of hyperspectral imagery classification and clustering technology are reviewed and summarized.Background and meaning.Secondly,the computing flow and evaluation criteria of hyperspectral imagery clustering analysis are introduced.Then some typical clustering algorithms are introduced.Then,based on the practical application environment of hyperspectral imagery,two points of improvement are proposed,Provided a theoretical basis.The clustering algorithm which is most suitable for the hyperspectral data environment is explored through the simulation experiment,and the traditional clustering algorithm is improved according to the improved idea,which proves that the pre-clustering algorithm can improve the clustering quality of the hyperspectral data.Thirdly,the algorithm flow and evaluation criteria of hyperspectral imagery classification and analysis are introduced.Several typical supervised and semi-supervised algorithms are introduced briefly.The hyperspectral imagery classification algorithm and twin support vector machine algorithm based on spatial information are emphatically introduced,which provides the theoretical basis and the original algorithm for the later algorithm improvement.Simulation results show that a large number of unlabeled samples are helpful for the classification of hyperspectral imagerys,and the cascading of spatial information andspectral information can verify the classification accuracy of hyperspectral imagery.Fourthly,the hyperspectral imagery spatio-spectrum semi-supervised classification algorithm based on clustering information is proposed.Firstly,the spatial information is extracted by 2D-Gabor filter,and the K_Medoids algorithm with improved depth is selected for the neighborhood selection of labeled samples.Then,the filtered samples are re-put into the dataset with the labeled samples using the idea of active learning Improve the classification accuracy of probabilistic model SVM.Finally,we proposed a semi-supervised cooperative classification algorithm of hyperspectral imagery based on fusion clustering features.Firstly,the clustering feature is extracted by the improved K_Means algorithm of hyperspectral data,and then combined with least squares twins support vector machine to form a collaborative training model.The model consists of clustering loss function,classification consistent function,classification difference,sample difference The four functions of the index formed by the objective function constraints.Combining the clustering features of collaborative training to make up for each other between the clustering and classification,effectively improve the semi-supervised classification accuracy.