Research On Hyperspectral Image Dimension Reduction And Classification Algorithms

In recent years,with the rapid development of remote sensing technology,hyperspectral remote sensing has gradually entered people’s field of vision and become a hot issue of research.At the same time,the related technology of imaging spectrometer is increasingly mature,which improves the quality of hyperspectral image acquired by imager.Therefore,hyperspectral remote sensing technology continues to improve.Hyperspectral images contain hundreds of narrow and continuous bands,which provide rich spectral information for the recognition and classification of objects.Meanwhile,Hyperspectral images have been widely used in various fields,including medical testing,environmental monitoring,marine development and so on.However,the large amount of data in hyperspectral,the considerable number of bands,and the relatively strong correlation between bands,pose a great challenge to their information extraction technology in recognition and classification.Thus,hyperspectral image dimensionality reduction and classification technology research have become the key issues of hyperspectral image processing,and relevant academic researchers have paid more attention on these two issues as well.In this paper,we mainly study the dimensionality reduction and semi-supervised classification of hyperspectral images.This dissertation summarizes the research achievements of scholars both at home and abroad.Aiming at the large amount of data,the large number of bands and the strong correlation between the bands in hyperspectral images,the Successive Projection algorithm is improved,which is applied to the reduction of the hyperspectral image processing.Considering the shortcoming of the traditional algorithm that the initial band is selected randomly,we introduce kurtosis and skewness value to limit the selection of the initial band.Therefore,it not only keeps the simplicity and speed of the original algorithm,but also improves the performance of dimensionality reduction.Experiments are carried out on the subsets of the bands obtained by dimensionality reduction experiments.It is found that the improved algorithm achieves the goal of efficient dimensionality reduction compared with the traditional algorithm.In order to solve the problems in the traditional M-training algorithm that small differences of the classifiers affect the classification accuracy of hyperspectral images when the number of labeled samples is not enough,this paper improves the standardM-training algorithm,Diversification,through the complementarity between different classifiers,enhancing the classifier performance.The classification error rate of unlabeled samples and the error rate of labeled samples are weighted,and each classification condition is set so that the labeled sample sets are effectively expanded.The experimental results show that the proposed algorithm outperforms the traditional M-training algorithm in terms of overall classification accuracy,average classification accuracy and Kappa coefficient.