Sparse Model Construction And Application In Remote Sensing High Performance Computing

With the development of network and sensor technologies,data from remote sensing area present the characteristics of diversified sources and huge volume.Traditional remote sensing processing methods have to face more challenges when dealing with the above huge and redundant data.Algorithms with more efficiency in knowledge discovering are highly demanded.As a new data mining technology,sparse representation roots in rigorous theories and has broad domain application verification.Concretely,by summing up its successful applications in natural image processing domain we can find its following merits: excellent latent structure discovery ability from complex data,powerful ability of feature finding and preservation,robustness of the image noises.Remote sensing image,one kind of natural images,itself has the natural sparsity which sparse representation methods are good at.Applicability and potentials of sparse algorithms for remote sensing image processing are needed to be further studied.Meanwhile,high real-time demanded processing tasks such as disaster monitoring challenge the image processing algorithms' performance and speed which only high performance computing can meet.However,implementation of high performance computing is very difficult,developers need to fully understand the target architecture and have ability of parallel programming.In fact,diversified hardware structures,complex parallel frameworks,optimal data allocation and reasonable resource scheduling are all challenges for remote sensing experts.In order to reduce the difficulties of parallel implementation,researchers from Niels Bohr Institute,University of Copenhagen,deployed a GPU cluster named Manjula and developed an auto-parallel tool named Bohrium.The combination of Manjula and Bohrium shields users from the diversity of underlying hardware and the complexity of parallelization.Users only need to focus on their own domain research.Parallelization and acceleration of algorithms can be realized automatically without any intervention.Concretely,this paper focuses the study on sparse representation application in remote sensing image processing,including priori knowledge characterization,model construction and optimization.We systemly summarized the theories of sparse representation,detailly studied two key problems of sparse models: sparse coding and dictionary learning and applied these models into two kinds of remote sensing images' three types of processing tasks,four algorithms are developed and experimented,specifically as follows: unsupervised double sparse dimension learning algorithm for hyperspectral image dimension reduction,context based sparse classification algorithm also for hyperspectral images with small training sample size,an improved sparse classification algorithm and a sparse pan-sharpening algorithm for GF-2 with high spatial resolution.Experimental results of the four algorithms can not only verify the applicability of sparse based method in remote sensing image processing tasks,but also show that this kind of technology is an effective alternative solution to large-scale data tasks.To further speed up,we introduce a new sparse oriented parallelization technology(CSR5 and Sp GEMM)to expend Bohrium and use the combination of Manjula and extended Bohrium to parallel the high time-consuming sparse coding algorithm.It presents a reference way for the fast implementation of high performance computing of remote sensing.