DCT Based Approaches For Hyperspectral Imagery Processing

The huge quantity of information and the high spectral resolution of hyperspectral imagery(HSI)present a challenge when carrying out traditional processing techniques such as classification.Dimensionality and noise reduction improves both accuracy and efficiency,while retaining essential information.This research work explores the unique capacity of Discrete Cosine Transform(DCT)to preserve signal energy in few low-frequency coefficients,thus reducing noise and computation time in hyperspectral imagery preprocessing.Therefore,in this work,three different approaches are proposed including,Spectral DCT-based approach for hyperspectral dimensionality reduction(SDCT-DR),DCT-based preprocessing approach for Independent Component Analysis(DCT-ICA)in hyperspectral data analysis,and Cascade Spectral-Spatial Discrete Cosine Transform framework(CSS-DCT)for hyperspectral imagery classification.In the first exploration,a dimensionality reduction approach based on the spectral DCT designed specifically for hyperspectral imagery is proposed.Thus,a spectral DCT is performed on each pixel vector,considered as 1-D discrete signal(spectral curve).Then,energy preservation criterion is performed to select the retained DCT coefficients as a reduced subspace.In order to evaluate this approach,three classification methods,Support Vector Machine(SVM),K-Nearest Neighbor(K-NN)and Fisher's linear discriminant analysis(LDA)are applied on the new reduced data.Principal Component Analysis(PCA),Fast Fourier Transform(FFT),Independent Component Analysis(ICA)and Minimum Noise Fraction(MNF)were also implemented and the results compared those from the proposed approach.These experiments demonstrate that the proposed approach provides a highly efficient means for feature extraction.In the second exploration,a novel approach that applies DCT as preprocessing for ICA in hyperspectral data analysis is proposed.Indeed,ICA is computational cost,and given the absence of specific criterion for components selection,constrains its application in high-dimension data analysis.To overcome these limits,our method exploits the DCT characteristic of retaining only the most useful information in few low frequency coefficients,thus reducing noise and computation time.Subsequently,ICA is applied to transform the inputs to components independent as possible for subsequent hyperspectral processing.In order to evaluate a novel approach,the reduced dataset using(1)ICA with the commonly used preprocessing techniques which is PCA,(2)ICA without preprocessing,(3)ICA with FFT and(4)ICA with DCT preprocessing are tested with SVM,K-NN and LDA classifiers on two real hyperspectral datasets.Experimental results in both instances demonstrate that dataset after our proposed combined DCT with ICA preprocessing method yields superior hyperspectral classification accuracy.In the third exploration,a novel hyperspectral classification framework based on cascade spectral-spatial Discrete Cosine Transform(DCT)and linear SVM is proposed.In this framework,a spectral DCT is performed first on the original hyperspectral image to obtain a spectral profile representation where the most significant information is concentrated in few low frequency components and the high frequency components represent a noisy data.Then,a spatial filtering is performed on the noisy spectral components to further extract the rest of useful information in the spatial dimension.For the spatial filtering phase,three different techniques are investigated including:the local spatial 2D-DCT,the global spatial 2D-DCT and,the Deep Convolutional Neural Network(DnCNN).Finally,an inverse spectral DCT is applied on all components including the filtered ones to obtain the final denoised hyperspectral image which is introduced to the SVM classifier.Experimental results on three datasets show that the proposed framework using the global spatial 2D-DCT for the spatial filtering outperforms several state-of-the-art hyperspectral classification methods,including SVM with Radial Basis Function kernel(SVM-RBF),the Radical Basis Function Neural Network(RBFNN),CNN,Edge-Preserving Filtering classifier(EPF),Optimal Deep Belief Networks(O_DBN),and O_DBN with Texture Feature Enhancement(O_DBN-TFE).The proposed DCT based approaches in this dissertation are validated on real datasets based on classification accuracy quantitative assessment;using the three traditional measurements including kappa coefficient,Average Accuracy(AA),and Overall accuracy(OA).On comparing the obtained results of the investigated approaches,with other state of-the-art methods,it is obvious that the proposed approaches outperform the other methods.Furthermore,the obtained results from the proposed classification framework are promising and outperform the recent sophisticated deep learning based hyperspectral classification approaches.