Research On GPU Parallel Optimization Of Hyperspectral Image Classification

At present, hyperspectral remote sensing has been widely applied in many fields of Earth science, thus becomes a new technique for many Remote sensing applications, such as earth observation, mapping, resource exploration, disaster investigation, military reconnaissance and so on. Hyperspectral image classification is an import content of hyperspectral image analysis. Due to the large amount of hyperspectral image data and high complexity of the algorithm, the execution efficiencys of existing serial classification algorithms are often low, and it is difficult to meet the real-time classification in hyperspectral image procesing. With the advance of high performance computing technology in recent years, GPU general-purpose computing technology obtained rapid development. Compared with CPU, GPU has the advantages of more powerful parallel computing capability and higher floating point processing ability, it provides a new effective way to accelerate the classification of hyperspectral image.The spatial-spectral classification method combine spectral and spatial information together, tend to have better classification accuracy. But the addition of spatial information, further increased the amount of the calculation of classification algorithm. How to improve the execution efficiency of the algorithm at the same time ensure the classification accuracy becomes a key problem needed to be solved in hyperspectral remote sensing information processing. After analyzed the principle of hyperspectral image classification and GPU parallel computing, this thesis studied the optimization of spatial-spectral hyperspectral image classification method based on GPU/CUDA architecture. The primary work is as follows:Firstly, the thesis designed a parallel optimization method based on GPU for spatial correlation regularized sparse representation classification(SCSRC) method. The method has effectively improved the effect of sparse representation classification by adding the spatial correlation constraints in sparse representation classification model. But it needs to search for dedicated atoms in the training dictionary for each test pixel, so it’s computational cost is very high. The thesis first analyzed the performance of the serial algorithm, and then used the task decomposition and memory access optimization method to optimize the computing intensive solving process. The experiment using real hyperspectral image verified the classification efficiency of the method.Secondly, based on GPU/CUDA architecture, a parallel optimize method was proposed for spatial-spectral kernel sparse representation classifier(SSKSRC). Kernel sparse representation classifier(KSRC) is suitable for nonlinear separable hyperspectral data. SSKSRC uses the neighboring filtering to describe the spatial correlation of neighboring pixels based on KSRC model and has better classification effect. In the parallel optimize method, the massive parallel threads are fully used to accelerate the calculation of kernel matrix and the solving process of classification model. And a memory access optimization strategy is designed to reduce the data interaction between CPU and GPU. The contrast with the serial and multi-core parallel algorithm on the CPU platform proved the validity and efficiency of the optimization method.Thirdly, a study of the optimization method of the spatial-spectral classifiers with sparse representation and MRFs on CPU-GPU heterogeneous platforms was carried out. MRF is an effective tool for the modeling of image spatial information, the classifier, combining the l1/2 regularized sparse representation and MRF-based spatial prior under Bayesian framework, can achieve better result. Based on the characteristics of the algorithm, the thesis made reasonable task allocation and storage optimization for the calculation process of the classifier which based on sparse representation and MRF-based spatial prior. The parallel method comprehensively utilize the logic control of CPU and parallel computing ability of GPU to increase the efficiency of the serial algorithm. Compared with the serial algorithm, the classification efficiency of the parallel method is very high.Finally, considering the real-time demand in hyperspectral remote sensing information processing, the thesis studied the optimization of sparse multinomial logistic regression (SMLR) classifier. A real-time parallel classification method was implemented based on the parallelization refactoring on the iterative process of the serial algorithm. Although this method can meet the real-time demand of remote sensing applications such as target detection, military reconnaissance and biochemical monitoring, but it only use spectral information for classification, the accuracy has certain promotion space. In order to strike a balance between high classification accuracy and the execution efficiency, the thesis further studied an effective classifier which based on weighed MRF and SMLR and designed the corresponding parallel optimization method. The experimental results using multiple real hyperspectral images show that the method can achieve high classification efficiency at the same time ensure high classification accuracy.