Hyperspectral Dimensionality Reduction Analysis Based On Kernel MNF Transformation

Hyperspectral images can synchronously obtain spatial information and spectral information of earth targets and have extremely high spectral resolution,which makes it possible to precisely identify the target classes.Hyperspectral images also know an important development in an increasing number of applications with in resources and environment,industrial production,as well as in biomedicine and national defense security.However,at the same time,hyperspectral data have relatively more bands and volume.The number of bands can reach up to hundreds,which produce challenges to analysis and application of hyperspectral data.Dimensionality reduction is considered to be an effective method to solve the above-mentioned challenges.Dimensionality reduction is an important part in the processing and application of hyperspectral images.Hyperspectral dimensionality reduction aims to simplify and optimize image features.It could effectively express high-dimensional data information and reduce data size greatly,and which enables rapid and precise extraction of target information.The common methods of dimensionality reduction are divided into two major categories: band selection and feature extraction.This paper mainly focuses on feature extraction.Feature extraction methods can be categorized as linear and nonlinear methods.The common linear feature extraction methods include principal components analysis(PCA),minimum noise fraction(MNF),independent component analysis(ICA),linear discriminant analysis(LDA)etc.The common nonlinear feature extraction methods mainly include manifold learning,graph-based and kernel function methods.This paper mainly concentrates on researching kernel MNF(KMNF)algorithm.KMNF is a nonlinear feature extraction method,while it failed to obtain the desired effect for hyperspectral images in practical application.The main reason is that noise estimation results are inaccurate.Based on this problem,this paper proposes three methods which include optimized kernel minimum noise fraction(OKMNF),image segmentation-based KMNF(KM-KMNF)and superpixel-based OKMNF dimensionality reduction and classification integrative algorithm(SP-OKMNG-SP).Based on real hyperspectral images,these new methods are compared with different advanced dimensionality reduction methods.Experimental results indicate that these new methods can extract nonlinear features effectively and get quality feature extraction results.The main achievements of this thesis are as follows:1.The thesis explains linear and nonlinear dimensionality reduction methods logically.These methods include PCA,MNF,(kernel PCA)KPCA and KMNF.This thesis analyzes advantages and disadvantages of these dimensionality reduction methods based on algorithms theories.PCA sorts components based on the descending order of image information content after transformation.But PCA cannot guarantee that the first few principal components have the highest image quality.MNF generates new components ordered by image quality.However,the above-mentioned methods could only extract linear features of hyperspectral data and was unable to find the nonlinear features.Although KPCA is a nonlinear dimensionality reduction method,it cannot also guarantee that the first few principal components have the highest image quality.KMNF solves the above problems,but it is inaccuracy in the calculation of noise and cannot provide desired results in real applications.Therefore,this thesis proposes an approach which considers spatial and spectral information to improve noise estimation results.Experimental results indicate that the new approach can get more accurate noise estimation results.2.This paper presents an OKMNF for feature extraction of hyperspectral imagery based on KMNF.KMNF cannot provide desired results in real applications because of estimating noise inaccurately.Therefore,to improve the effect of KMNF,this paper proposes OKMNF which considers more accurate SSDC1 and SSDC2 noise estimation methods to estimate noise in KMNF.Based on two real hyperspectral images,Experimental results indicate that the proposed OKMNF benefits significant improvements in dimensionality reduction.3.Based on OKMNF,this paper proposes an image segmentation-based KM-KMNF method.KM-KMNF is an advanced optimal OKMNF.Image segmentation is used to divide image into several subregions,which do not intersect each other.The spectral features are homogeneous within each subregion.The proposed method exploits the homogeneous region generated by image segmentation as the minimum region for multiple linear regressions.By integrating the spatial information of the homogeneous region with spectral decorrelation,the proposed KM–KMNF improves noise estimation for feature extraction,thus enables better performances on feature extraction.4.Based on all above research,this paper proposes a superpixel-based OKMNF dimensionality reduction and classification integrative SP-OKMNF-SP algorithm.The emphasis of SP-OKMNF-SP can be divided into two parts.Firstly,in dimensionality reduction part,SP-OKMNF-SP can be more effective promote the performance of dimensionality reduction for hyperspectral data by introducing the spatial information of superpixel to improve the precision of noise fraction.Besides that,in image classification part,SP-OKMNF-SP performs image classification based on superpixel as the basic unit instead of operating at the pixel level.Based on two real hyperspectral images,Experimental results indicate that the proposed SP-OKMNF-SP benefits significant improvements in hyperspectral dimensionality reduction and image classification.