Research On Hyperspectral Image Dimension Reduction And Semi-supervised Classification Algorithms

At present,hyperspectral remote sensing is the frontier of remote sensing.With the development and maturity of the imaging spectrometer,hyperspectral remote sensing has been developed rapidly.As a result,the level of acquisition of hyperspectral images is gradually increasing.The obtained hyperspectral image contains hundreds of narrow and continuous bands,providing abundant and subtle spectral information for identifying and classifying of objects.Hyperspectral images have been widely used in medical imaging,earth monitoring,resource exploration,urban environmental testing,etc.However,the large amount of hyperspectral data,the considerable number of bands,and the relatively strong correlation between each other,owing to these characteristics,it is a great challenge to the information extraction technology for the identification and classification.For hyperspectral image processing,the research of hyperspectral image processing and classification technology have become the key issue,and they has been widely concerned and attentioned by researchers.In this paper,the dimensionality reduction and semi-supervised classification of hyperspectral images are studied.In this paper,we summarize the research achievements of scholars at home and abroad.Due to the large amount of data in the hyperspectral image,the large number of bands and the relatively strong correlation among the bands.The fireworks algorithm is used to reduce the hyperspectral data,Compactness-Separation Coefficient is used as the measurement criterion.Compared with the genetic algorithm and the tabu search algorithm,the results of simulation experiments show that the proposed algorithm has achieved better performance in reducing the number of bands,overall classification accuracy,Kappa coefficient and running time.When the standard Tri-training algorithm has a small number of label samples,there is no obvious difference between the classifiers;this phenomenon affects the classification accuracy of hyperspectral images.In this paper,the standard Tri-training algorithm is improved,and the semi-supervised classification of the Tri-training algorithm with enhanced difference is proposed.The support vector machine,random forest and nearest neighbor are introduced into the standard Tri-training algorithm as base classifiers,using the cooperative and complementary relationship between different classifiers.At the same time,stratified sampling based on labeled class is used to deal with the over-labeling of the unlabeled sample.This strategy is better to ensure that the subset of samples extracted is representative.The results of simulation experiments show that the proposed algorithm outperforms the standard Tri-training algorithm in overall classification accuracy,average classification accuracy and Kappa coefficient.