Research On Classification-oriented Feature Extraction From Hyperspectral Image

The characteristics of high spectral resolution and numerous bands of hyperspectral remotesensing image has brought great convenience to fine recognition features and causes the dramaticincrease in the amount of image data.When use all bands into information extraction andclassification of treatment,not only lead computer processing load increase substantially andinefficiency,but also hughes phenomenon can be happened.This paper are based on the analysisof hypersepctral image classification and feature extraction,and make full use of hyperspectralremote sensing technology to get a good balance between precision,efficiency and methods.Thispaper gave a summary of feature extraction and classification methods of hyperspectralimage,several researches on classification-oriented hyperspectral image feature extraction weremade.Some achievements have been made as follows:(1) From the describing space of hyperspectral image on,the essentiality of dimensionreduction was given from the aspect of classification.Then the feature extracion andclassification technologies were investigated,and the potential of classification-orientedhyperspectral image feature extraction was expatiated.(2) Several methods of Intrinsic dimensions were summed up,Maximum likelihoodestimator was analysed,because the method global convergence is poor, ignoring the contributionof a single sample point,this leads to important information is the annihilation of thedefects,redundant information increases.In the paper,Adaptive maximum likehood estimator isproposed to solve the problems.the use of adaptive maximum likelihood estimation,byexperimental verification of the continuation of the advantages of maximum likelihoodestimation method not only in the theory and calculations,but also take full account of thedistruibution of the data,above the maximum likelihood estimate of the intrinsic dimensionaccuracty estimation method,and this method is lessensitive to changes in the number ofneighbors.(3) To improve the classification’s result of the Locally Linear Embedding(LLE),solve theconfusing phenomenon,the reconstruction error is used to classify the test sample and extensionfactor to verfy object’s distacne.To verfy the effectivenss of the proposed method,theexperiment is conducted on hyperspectral imagery’classification and the experment resultshows the improvement method can significantly improve the classification accuracy.(4) In the paper,a feature extraction on Two dimensiona principal Component Analysis areused on the field of Hyperspectral image feature extraction.This method can make full use ofdata’space information,it can find the best project’direction by linear transformation.Thismethod can make the object with the same lable compact together,separate the object with different lable.At the same time,it can elimate the final results of the “pitting” phenomen andavoid classification surface features confusion,the result of classification can be smoothed by2DPCA.