Research On Parallelization Algorithm Of Hyperspectral Image Unmixing

Spectral imagers in the electromagnetic field of ultraviolet,visible,near infrared and mid-infrared region,with hundreds of continuous and subdivided bands on the target area at the same time imaging,get hyperspectral images.In practice,the process of hyperspectral image acquisition is affected by many aspects,such as shooting distance,material interference in the atmosphere,and resolution of the spectrometer lens.On the other hand,the target area contains the uncertainty of the object As well as the randomness of the distribution and blending of the objects,causing the pixels contained in the image to be mixed pixels.The existence of mixed pixels leads to the uncertainty of the feature information.Therefore,hyperspectral image unmixing is an essential research.In this paper,the study of hyperspectral image unmixing is based on a linear hybrid model.In this paper,the hyperspectral image decoupling study using the theory of blind separation theory,the abundance value and one,the abundance of non-negative as the objective function of the solution.In this paper,we propose a differential search algorithm based on elite strategy and probability selection.The improved algorithm can quickly converge to the complex function in solving the problem.In this paper,the improved algorithm can solve the problem.In this paper,the objective function is solved by the differential search algorithm based on elite strategy and probability selection,and the efficiency of the whole algorithm is improved,which makes the optimization process fast and stable.In this paper,two parallelization strategies are proposed,which are based on the hyperspectral image solution.In this paper,we propose a new method to solve the problem of hyperspectral image unmixing.Mixed pretreatment and intelligent optimization phase,and described their work model and implementation process respectively.In this paper,hyperspectral simulation data and real remote sensing data are used to evaluate the performance of the algorithm.The experimental results show that the proposed solution can achieve high spectral solution.The main work of the thesis is as follows:(1)Optimization of the new algorithm differential search algorithm,differential search algorithm is a new and efficient group of intelligent algorithms.But the algorithm still has the disadvantage of slow convergence and high precision.To this end,this paper presents an improved method.Firstly,the optimal value of the population is introduced into the search equation to guide the first search,so that the algorithm can reach the global convergence quickly.On the other hand,the selection probability P is used to guide the search equation,which improves the convergence speed and the search precision.(2)The objective function of the independent component analysis of the constraint condition is easy to fall into the local convergence due to the influence of the initial setting and the iterative step size.The objective function is solved by the difference search algorithm based on the elite strategy and the probability selection.To achieve non-negative signal blind separation,that is,to complete the mixed pixel solution.(3)Because the hyperspectral image unmixing is vast,the data is huge,in the process of dealing with a large number of calculations,consuming a lot of time.In this paper,we study the parallelization of hyperspectral solution based on CUDA.According to the characteristics of hyperspectral solution,two parallel schemes are proposed.