An Adaptive Parameter Method For Hyperspectral Image Classification Based On Multi-GPU

Hyperspectral image classification is one of the most important research topics of remote sensing image processing.The rich spectral information and spatial information contained in hyperspectral remote sensing images can clearly reflect the unique spectral characteristics of features,thus it has significant advantages in accurate feature recognition.Hyperspectral remote sensing image classification been successfully applied in geological research,biological sciences,ocean exploration,atmospheric monitoring and other fields,which is of high research significance and application value.However,with the continuous development of hyperspectral imaging technology and classification algorithms,the amount of data in hyperspectral images is becoming larger and the algorithm complexity is getting higher.The traditional computer and serial processing algorithms can no longer meet the needs of rapid classification of hyperspectral images.Therefore,how to achieve fast and efficient hyperspectral remote sensing image classification has become one of the key issues in the development of hyperspectral remote sensing classification technology.With the development of high-performance computing technology,GPU is widely used in various scientific computing fields due to its powerful parallel computing capabilities,providing new opportunities for efficient classification of hyperspectral remote sensing images.In order to improve the time performance of hyperspectral image classification,based on the research of CPU+multi-GPU heterogeneous computing technology and CUDA parallel programming,this paper deeply analyzes the spacial spectral kernel sparse hyperspectral image classification method,and proposes a multi-GPU Parallel optimization algorithm for hyperspectral classification.At the same time,in order to further improve efficiency and stability,a multi-objective optimization load balancing optimization algorithm based on cuckoo search algorithm is proposed.Experiments have proved that the multi-GPU parallel classification algorithm and load balancing algorithm proposed in this paper can make full use of the computing resources of heterogeneous systems,improve classification efficiency,and obtain a better pareto frontier under the premise of ensuring the accuracy of algorithm classification.The main work of this paper include:（1）Designed a multi-GPU parallel algorithm of spatial-spectial kernel sparse representation for hyperspectral image classification.Based on analyzing the spatial-spectial kernel sparse representation classification method and its data dependence and computational parallelism,this paper reconstructed the algorithm calculation process and designed the multiGPU parallel implementation.Besides,this paper designed an optimized storage mode to improve GPU occupancy.Experiment results show that the multi-GPU parallel optimization method proposed in this paper can achieve efficient and real-time hyperspectral remote sensing image classification on a CPU+multi-GPU heterogeneous system under the premise of ensuring the classification accuracy,with a maximum speedup of 55.15 x.At the same time,this paper proposed an adaptive optimization method of GPU kernel parameters for hyperspectral classification and multi-GPU parallel optimization.Based on analyzing GPU structure characteristics,CUDA programming,the influence of CUDA kernel function parameters on grid and warp scheduling deeply,combined with the allocation of GPU resources,an adaptive optimization calculation method for kernel parameters is proposed.Experiments show that the kernel parameter adaptive optimization method proposed in this paper can effectively guide the adaptive configuration on the GPU,and further improve the computing efficiency and GPU computing resource utilization.（2）Designed a multi-objective optimization load balancing model for hyperspectral image classification is.Based on the detailed analysis of the parallelism of spatial-spectial kernel sparse representation classification algorithm,this paper designed a reasonable task division mode on the CPU+multi-GPU heterogeneous platform,and proposed a multi-objective optimization load balancing model for hyperspectral image classification.Furthermore,to solve the model,this paper designed a heuristic algorithm based on the cuckoo search algorithm.The experimental results show that the designed method can obtain a better pareto frontier,and further improve the system efficiency and stability.