Research On Spectral-spatial Feature Mining For Hyperspectral Image Classification

Hyperspectral imagery(HSI)classification is one of the major research topics in the field of image processing and interpretation.The spatial and spectral resolutions have been increased with the recent development of hyperspectral remote sensing technology.In the meanwhile,new challenges are brought to HSI classification.The major issues are:(1)High spectral dimensionality and data redundancy;(2)Complex nonlinear characteristics;(3)Lack of training samples;(4)“Same spectral information,but different class”,“different spectral information,but same class”;specifically,issue(4)will result in low classification accuracy since the classification algorithms are based on spectral information.Additionally,the spectral redundancy and lack of training samples will bring about “Hughes phenomenon”.Therefore,the major focus of this dissertation is to investigate how to fully exploit the spectral and spatial information and how to develop efficient classification frameworks which will improve the classification accuracy.The major aspects of this dissertation are as follows:1.High dimensionality and nonlinear separability are two important properties of HSI;however,the traditional spectral-spatial shared ridge regression(SSSRR)classifies in the linear space and ignores the non-linear separability between various objects for HSI.To overcome this drawback,this dissertation proposes Spectral-spatial shared kernel ridge regression(SSSKRR),which uses the kernel ridge regression(KRR)and nonlinear sharing subspace method to improve the traditional SSSRR.The datasets has larger separability with SSSKRR,resulting in higher classification accuracy compared with the linear approach.Moreover,an efficient iterative optimization algorithm based on Singular value decomposition(SVD)is used to solve the objective function of SSSKRR,thus avoiding the high time complexity in the traditional kernel learning algorithms.Experimental results on several real hyperspectral datasets have shown that the proposed approach in this chapter is fast,robust,and accurate for hyperspectral imagery classification.2.Lack of training samples is one of the main factors that affect the classification accuracy of HSI classification,and the Support Vector Machine(SVM)can have satisfactory classification performance for small training sample datasets.Traditional spatial-spectral SVM-based algorithms for HSI classification adopt unsupervised algorithms such as Markov Random Field(MRF)and Graph Cut(GC)to adjust the output labels of SVM.Those algorithms do not exploit the label information fully and may lose some useful information for classification.To tackle this problem,in this chapter,Spatial Logistic Regression(SLR)for Support Vector Classification(SVM+SLR)is proposed for HSI.The proposed algorithm is a two layer learning model,which uses the spectral and spatial information by layers.First,the original hyperspectral data is processed by spectral-based SVM to obtain the onedimensional feature.Then the neighborhood information of the dataset is extracted by a rectangular window.At last,the spatial information is classified using a supervised logistic regression algorithm.Experimental results show that the proposed approach can significantly reduce the testing time while effectively improving the classification accuracy for HSI.3.When the number of training samples is extremely small,the classification accuracy of SVM+SLR is still low because SVM+SLR used linear based logistic regression model.In order to solve this problem,a Spectral-spatial SVM-based multi-layer learning algorithm(SSMLL)is further proposed for HSI classification.SSMLL is a three layer learning model.First,spectral-based SVM is used to process the original HSI,and one dimensional feature is obtained.Second,sigmoid function is used to scale the SVM output and enhance the nonlinear structure.Third,rectangular window is used to extract the neighborhood information,and finally,the spatial information is further classified using the kernel SVM.The nonlinear characteristic of the third layer kernel SVM is very helpful to improve the classification accuracy.Experimental results show that the proposed SSMLL have great potential in HSI classification applications.4.Although the single local neighborhood feature based approach can obtain relatively high classification accuracy,the single spatial feature can only describe one characteristic of hyperspectral imagery,and thus,the classification performance can be improved.To solve this problem,a multiple feature-fusion and kernel principle component analysis(KPCA)SVM based classification method is proposed for HSI classification.Extended Multi-attribute profiles(EMAP)and local binary pattern(LBP)are two efficient spatial features.The proposed MF-KPCA first extracts the EMP and LBP features of HSI,and then KPCA is conducted to reduce the dimensionality of the three features,including EMAP features,LBP features,and the original spectral features.Finally,the features are stacked,and SVM is used to obtain the classification accuracy.EMAP and KPCA are first introduced to HSI simultaneously,and KPCA can not only reduce the redundancy information to reduce the computational complexity,but also obtain more separable features to improve the classification accuracy.5.Traditional EMAP feature based algorithms use PCA to reduce the dimensionality of each principle component and extract the EMAP feature of each principle component.However,the feature map obtained by the traditional PCA cannot preserve the spatial structure,and the classification accuracy can be affected.In addition,the EMAP features have a lot of redundant information,and noises can affect the performance.To solve this problem,this dissertation proposes an EMAP,image fusion and recursive filters based SVM classification method(EMAP-IFRF).EMAP-IFRF first uses an unsupervised band selection algorithm linear prediction error(LPE)to replace the traditional PCA and extracts the EMAP features of each selected band;then,image fusion(IF)is used to reduce the dimensionality of the EMAP features.Finally,domain transform recursive filter(RF)is used to smooth the EMAP feature map while preserving its edge structure,and SVM is used for classification.LPE and RF can obtain better spatial features to improve the classification accuracy.Additionally,the proposed method has much less memory requirement and training time due to the reason IF is used to reduce the dimensionality of the EMAP features.