Hyperspectral Spectral-spatial Joint Classification Method Based On Orthogonal Discriminant Analysis

With the rapid development of hyperspectral remote sensing technology,the imaging spectrometer can capture the precise spectral response and spatial details of surface objects,which makes that the hyperspectral remote sensing image are rich in spatial information and spectral information.However,the rich spectral information of the hyperspectral remote sensing image means that the number of bands is large,resulting in high redundancy and correlation among bands.Therefore,the hyperspectral remote sensing image provide abundant information and bring great challenges for subsequent classification applications.How to effectively mine and utilize the rich information of the hyperspectral remote sensing image to improve the classification effect is a key issue,which has received extensive attention.This article attempts to unify the three aspects of feature extraction of spectral dimensions,spatial filtering of three-dimensional structures,and utilization of spatial information up and down to a framework,fully exploit spatial information and spectral information of the hyperspectral remote sensing image,which improve the classification accuracy of the hyperspectral remote sensing image.This paper presents a classification algorithm based on orthogonal linear discriminant analysis and the markov random field.Firstly,the hyperspectral remote sensing image with less components removed by principal component analysis are performed by orthogonal linear discriminant analysis to achieve features extraction and features reduction;secondly,the extracted features are decomposed from different scales,frequencies and directions by three-dimensional discrete wavelet,which obtain the spatial-spectral fusion features set with orthogonal class discriminant information in a cascaded manner;then the spatial-spectral fusion features set can become input features of the probabilistic support vector machine,and markov random field theory is introduced in the classification process to establish the local spatial-temporal consistency of adjacent pixels,which can use spatial context information to correct the initial classification map to obtain the classification result.The experimental results on two real data sets show that compared with the existing clas sification algorithms,the proposed algorithm can achieve higher overall classification accuracy and Kappa coefficient,and effectively improve “salt-and-pepper noise”phenomenon and wrong points in the classification.