Hyperspectral Image Classification Using Kernel Joint Sparse Representation And GPU Implementation

Since hyperspectral remote sensing combines the spectral technology with imaging technology,it can accurately identify and classify the ground objects by extracting their effective features.Imaging spectrometer can utilize many adjacent electromagnetic wave bands to obtain hundreds of narrow bands of spectral information,such that each pixel is a continuous spectral vector.Although hyperspectral image(HSI)has the advantages of high resolution and the properties such as multi-band,image and spectrum,the noises and the high dimensions often make it more difficult to classify for HSI.In addition,due to the large volume of data,using the single-thread computational processing by traditional central processing unit(CPU)is difficult to achieve the rapid classification in the practical application.In view of this,by considering the characteristics of HSI to address the above problems,the model that based on sparse representation theory is studied to improve the classification performance.Moreover,the graphics processing unit(GPU)-based high-performance computing is adopted to implement the parallelization of the algorithm.The main research content of this thesis includes the following three points:1.Starting from the theory of sparse representation,this thesis first studies the current mainstream HSI classification algorithms,e.g.,sparse representation classification(SRC)and joint sparse representation classification(JSRC).Then,their coefficient solving methods such as orthogonal matching pursuit and simultaneous orthogonal matching pursuit also are reviewed.More importantly,by considering the importance of spatial information for HSI classification and the problem that JSRC classifier does not consider the contributions of neighbors for testing pixels,this thesis uses Gaussian kernel function-based similarity distance to calculate the similarity between local pixels and assigns weights to the neighbor pixels within the sliding window.By establishing this weighting mechanism,the classification model makes more reasonable use of spatial information and performs better classification performance.2.The traditional sparse representation is linear classification method which does not consider nonlinear characteristics of HSI.To address the above problem,this thesis introduces the idea of kernel mapping.The weighted joint sparse representation model is extended to Gaussian kernel space and a weighted kernel joint sparse representation classification(WKJSRC)model is proposed.Two real hyperspectral images are utilized for experiment and the experimental results show that the WKJSRC model obtains better classification performance than that of others.3.Besides the HSI has a complicated structure and huge volume of data,the incorporation of spatial information also further improves the complexity of the algorithm.Thus,using traditional CPU to process HSI data is inefficient.From this point of view,the thesis uses CUDA-C language based on the NVIDIA CUDA framework and CPU/GPU cooperative computing to parallelize the proposed WKJSRC algorithm.By making full use of the powerful parallel computing capabilities of GPU to implement the parallelization of high-density and parallel-processing operations in algorithm,a parallel algorithm of WKJSRC model is implemented.Thus the computational efficiency is improved.