Research On Sparse Blind Unmixing Methods For Hyperspectral Images Based On Particle Swarm Optimization

Hyperspectral remote sensing images can simultaneously provide spatial distribution and rich spectral information of ground covers and have high application value in many fields,such as national construction.However,a large number of mixed pixels seriously affect the accuracy of hyperspectral remote sensing applications.Therefore,spectral unmixing technology is used to extract finer ground information from hyperspectral images to obtain spectral characteristics of each material that make up each pixel and its corresponding abundance ratio.Most of the pixels of hyperspectral images are generally only composed of a small number of endmembers,and the overall spatial distribution of each material has obvious sparsity.It is of great theoretical and practical significance to carry out high-precision sparse blind unmixing research to estimate endmember and abundance simultaneously.There are three main problems with the traditional sparse blind unmixing methods.1)Gradient optimization methods rely on a one-way search to estimate the unknown variable of non-convex sparse blind unmixing problems,which are easy to fall into local optimum.2)The expression of abundance sparsity underutilized the spatial characteristics of the image.3)It is sensitive to the setting of the penalty coefficient of sparse terms.On the other hand,the swarm intelligence algorithm represented by particle swarm optimization has robust global search performance through the competition and cooperation between different individuals in the swarm,making it easy to find accurate solutions to all kinds of complex mathematical optimization problems.Therefore,this paper designs and implements comprehensive learning,spatial reweighted,and adaptive penalty mechanisms under the framework of particle swarm optimization algorithm,proposes three new sparse blind unmixing methods and verifies the superiority of the proposed method over other classical and advanced sparse unmixing methods through comprehensive experiments of simulated data and real hyperspectral images.The main research contents of this paper are as follows:1.Given the problem that gradient optimization methods are prone to local optimization and particle swarm algorithms’ premature convergence in dealing with high-dimensional issues,a hyperspectral image sparse unmixing approach based on comprehensive learning particle swarm optimization is proposed.This method first adopts the dimension division of fitness function and the construction of endmember and abundance double swarm to improve the accuracy of information exchange between particles.Then,two comprehensive learning strategies are designed to further refine the local position search of particles in the problem subspace and learn the global abundance sparsity of each endmember,to improve the performance of sparse unmixing.2.Aiming at the problem that expression of abundance sparsity fails to exploit sufficiently local and global spatial information of images,a multilevel reweighted sparse unmixing method based on superpixel segmentation and particle swarm optimization is proposed.Firstly,superpixel segmentation is performed on hyperspectral images,and two local reweighted factors are constructed according to the sparseness and the low-rank property of pixels’ abundances in the generated superpixels.Then,according to the sparseness changes of each material abundance map,a global sparse weighted factor is constructed.The traditional sparse unmixing mathematical model is reconstructed with three weighted factors,which reasonably expresses the sparsity of material abundance at different spatial levels.Finally,the particle swarm optimization framework is employed to achieve the solution to improve the accuracy and reliability of the unmixing.3.Given the sensitivity of traditional methods to the setting of sparse term penalty coefficients,a hyperspectral image sparse unmixing method based on adaptive penalty particle swarm optimization is proposed.This method organically combines the dynamic information about the feasible solution ratio of particle swarms based on unmixing reconstruction error within the dual swarms’ particle swarm optimization framework based on dimensional division and further designs a new adaptive sparse penalty mechanism.In the process of unmixing,the influence of data reconstruction error and sparse regular terms on search performance is effectively weighed in sparse unmixing,which can effectively improve the unmixing accuracy while overcoming the sensitivity problem of the penalty coefficient.