Research On Parallel Optimization Method Of Tensor Computation For Remote Sensing Image Fusio

Remote sensing images have been used in many domains,including ecological environment,vegetation characteristics,and military inspections because they contain rich feature information.Tensor can retain the structural constraints of remote sensing images.Therefore,it is of great scientific value to represent the remote sensing image by tensor.Fusion of remote sensing image can enrich the spectral and spatial data characteristics of images.At present,the remote sensing image fusion algorithm based on tensor has significant advantages in fusion effect.However,due to the large volume of remote sensing images and the high computing requirements of the fusion algorithm,the traditional single-machine processing mode is not applicable to the processing of remote sensing data fusion based on tensor calculation.Nevertheless,using cloud computing technology to optimize the remote sensing image processing in parallel can effectively solve the bottleneck problem in the case of single machine.A series of parallel optimization methods based on HDFS and Spark framework are proposed in this thesis.The methods are designed to address data storage and computational efficiency problem of basic t-product calculations,tensor image denoise and remote sensing image fusion algorithm.The main contents of this paper are as follows:(1)Based on the Spark platform,this paper proposes a parallel method of seven basic tproduct calculations:T-FFT,T-IFFT,T-product,T-SVD,TNN,TSN,T-SVT.To achieve better performance,strategies of reducing transmission variable communication consumption,dividing tensors properly,and avoiding data skew are proposed.The experimental results show that the above parallel optimization algorithm could improve the computational efficiency while ensuring the computational accuracy and the speedup can be up to 15.26.In order to improve the reuse of the t-product parallel algorithms,the above algorithms are packaged into the tensor product operation parallel toolkit SP_Tproduct,which provides standardized function interfaces,realizes the separation of call and execution.(2)In order to improve the computational efficiency of TRPCA,a tensor-based image denoise algorithm,this paper proposes a parallel optimization method SP_TRPCA.The method combines the function of SP_Tproduct with Spark platform.To make full use of Spark memorybased features,strategies of caching RDDs with high usage rates is implemented.Experiment results show that SP TRPCA has a good denoise effect and reaches speedup of 9.98 maximumly.(3)In order to improve the computational efficiency of NPTSR,a tensor-based remote sensing image fusion algorithm,this paper proposes a parallel optimization method SP_NPTSR.We develop the parallel method of remote sensing tensor patches extraction and complex tensor iteration calculation.The parallel remote sensing tensor patches extraction method includes strategies of restoring tensor,dividing tensors properly and designing the parallel algorithm.Meanwhile,the parallel complex tensor iteration calculation method optimizes the complex matrix multiplication.In addition,a series of optimization strategies such as reducing data transmission across partitions are implemented.Through experimental comparisons under different data sets and different degrees of parallelism,it is verified that SP_NPTSR improves the execution speed of the algorithm while maintaining the fusion accuracy.(4)Based on the research above,an efficient parallel processing platform for remote sensing images that supports tensor calculation is designed and implemented.Through componentized tensor parallel algorithms and customized remote sensing image processing procedures,users can use online cloud computing resources on-demand.Thus,the complex parallel processing of remote sensing images is completed.In conclusion,the platform has good scalability and practicability.