GPU-based Hyperspectral Image Classification And Target Detection

With the rapid development of hyperspectral remote sensing technology,it has been playing a very important role in Earth science,such as image recognition and classification,mapping,disaster investigation and prevention.Nowadays,hyperspectral image data has been so large and efficiency of the existing algorithms in this fields is too low to meet the requirement of the real-time in applying.As time going on,GPU fields has been growing rapidly.With the drive of NVIDA company,CPU/GPU heterogeneous platform based on CUDA has drawn much attention from researchers.This platform offers a good way to solve above problem for its high parallel computing ability and high logic operation ability.This paper bases on CPU/GPU heterogeneous platform and combines three algorithms of hyperspectral image field.Through the real data and experiment,this paper demonstrates that the method of GPU/CUDA can dramatically decrease the time cost of original algorithm while keeping the same algorithm result.The main content of this paper is as follow:Firstly,a parallel optimizing method has been designed for hyperspectral image kernel sparse representation classification based on spatial-spectral graph regularization.Based on KSRC model,Spatial-Spectral Weighted Graph and Sparsity Concentration Index have been adopted into this model to form the new method,called SWGSCI-KSRC.Through the real experiment,the SWGSCI-KSRC method has been proved its high classification efficiency and robustness.However,it takes one hour when it deals with the data of 100MB.In order to enhance the efficency,our paper analyzes the SWGSCI-KSRC algorithm and takes the strategy of share memory and global memory into consideration,makes some strategies in task allocation as well as taking full advantage of kernel function in CUDA.Finally,the parallel experiment dramatically decreases time cost,gets the 41 speedup while keeping the same classification result.Secondly,a parallel optimizing method has been designed for classification of hyperspectral image based on LRDSS.In the parallel experiment of this paper,we uses the strong logic operation ability of CPU and high parallel computing ability of GPU to enhance the algorithm efficiency.We fully make use of hardware structure,make balance between CPU and GPU,decrease the cost of interaction and finally get a high speedup result.Thirdly,a parallel optimizing method has been designed for hyperspectral images anomaly detection based on LRASR.Based on LRASR algorithm,the paper fully uses the GPU feature of high bandwidth,strong capability,low cost and highly cost effective.When interaction occurs,we take bank conflict into consideration and use texture memory to enhance the local efficiency in order to achieve fine-grained parallel.Also,SVD can be finished by library function after optimization.Finally,the parallel experiment dramatically decreases time while keeping the same algorithm result.Fourth,Visual Studio/MFC is used for designing HICTDS.The main module includes system frame,process design,the display of hyperspectral image,the operation of spectral library,spatial-spectral classification,target detection as well as their realization and test analysis..