GPU Parallel Classification Of Hyperspectral Images Based On Generalized Combined Kernel

Hyperspectral remote sensing technology has been widely applied in the earth sciences,such as earth observation,resource exploration,mapping,military reconnaissance,disaster investigation and so on.It is a new technique for many Remote sensing applications.Hyperspectral image classification is one of the hottest research fields of the hyperspectral remote sensing area.Due to the large-scale of hyperspectral data and the computational complexity in hyperspectral image classification,furthermore,the execution efficiency of the normal serial classification algorithms are often low based on CPU and it is difficult to meet the demands of real-time classification in hyperspectral image processing.Recent years,the GPU general computing technology has been rapidly enhanced.Compared to CPU,it has the advantages of multi-GPU processing cores and strong processing power and high memory bandwidth.GPU has the potential of accelerate the efficiency of large data processing for high-performance computing technology.In this paper,based on the analysis of generalized composite kernel hyperspectral image classification model and algorithm,a hyperspectral remote sensing image system for real-time applications is designed and a real-time parallel classification method based on CUDA is proposed.Finally,a generalized composite kernel framework hyperspectral image classification system based on GPU has been designed using MFC,and the experimental test and analysis.Papers mainly include:First of all,aiming at the classification system of hyperspectral remote sensing images,the theory,principles and user requirements of hyperspectral remote sensing image classification for real-time applications are deeply analyzed.A method of classification of hyperspectral remote sensing images for real-time applications is designed.Based on this,a GPU-based hyperspectral remote sensing image parallel classification system is designed.Secondly,on the basis of analyzing the generalized composite kernel hyperspectral image classification model and algorithm.Based on CUDA/CULA,this paper designs corresponding parallel optimization for generalized combinatorial kernel,LORSAL classification algorithm and polynomial logistic regression.Hyperspectral image classification of generalized composite kernel functions has great flexibility in the combination of kernel functions,and the method of convex combination is abandoned.And this method is to model the spatial information by exploiting the extended multivariate feature morphology.At the same time,we can also control its generalization ability through logistic regression.Finally,based on CUDA hyperspectral real-time parallel classification,a generalized composite kernel architecture hyperspectral classification system based on GPU was designed by using MFC.The system framework designed in this paper mainly includes the following modules:hyperspectral image data selection,parameter setting,classification processing and image display and speedup calculation.Finally,the real-scene hyperspectral image data is used to verify the parallel classification system proposed in this paper.