Particle Swarm Optimization Based Hyperspectral Image Endmember Extraction

Hyperspectral imaging technology is the foreland of the remote sensing and is one of the most important focuses of the remote sensing domain.Imaging spectroscopy allows for the synchronous acquisition of both images and spectra of objects,which provide special information for us to observe the earth.However,mixed pixels often exist in the hyperspectral image(HSI)due to the limitation of the spatial resolution of sensors and the complex distribution of surface objects,which have hindered the high precision interpretation of the HSI.How to effectively interpret the mixed pixel is a key problem in remote sensing applications.Endmember extraction is one of the most important methods to solve the mixed pixel problem.Endmember extraction can obtain the constitute materials of mixed pixels,whichdetermines the accuracy of the image interpretation such as spectral unmixing and target detection.Therefore,how to extract the complete and most representative endmembers has become one of the key and hot issues of the HSI analysis.There are several challenges in endmember extraction:1)the real HSI is usually corrupted by noise or outliers,where the data structure in the feature space can notmeet the convex assumption that traditional methods usually based on;2)due to the complexity of the HSI,the endmember extraction results are different when using objective functions that describe different characteristics of the HSI;3)spectral variability may exist in the HSI caused by the grain-size distribution of minerals or variations in illumination.In this thesis,several strategies are designed according to the characteristics of endmembers,and the particle swarm optimization are used to optimize the objective functions because of its advantage of advantages of being simple and demanding few parameters.The main contents of this thesis are as follows:(1)A novel endmember extraction method is developed,which is based on quantum-behaved particle swarm optimization(QPSO).The endmember extraction problem is described as a combinatorial optimization problem with minimizing the root-mean-square error as the objective function,avoiding the dependence on data structure.The performance of the particle swarm optimization(PSO)is limited due to its premature convergence,and much effort is required to enhance the optimization result.To address this problem,a novel QPSO algorithm is proposed for hyperspectral endmember extraction.A row-column coding approach for the particles is designed by combining the characteristics of quantum particles and the smoothness characteristics of HSIs.Experimental results show that this coding method is more efficient than the pixel-number based coding method.A cooperative method is employed to conquer the"curse of dimensionality" problem according to the searching mechanism of particles.This approach splits the multi-dimensional particles into one-dimensional subparts when updating the individual best position and the global best position of the particle.The cooperative method can improve the precision of endmember extraction in the multidimensional case.(2)An endmember extraction algorithm based on multi-objective particle swarm optimization(PSO)is proposed.The motivation and necessity of multi-objective endmember extraction is analyzed by taking the most commonly used two objective functions,the volume maximization and the root mean square error minimization as examples.For the multi-objective optimization problem,different optimization functions often conflict with each other during the process of optimization,which means that finding a solution that optimizes all the objective functions at the same time is almost impossible during the process of optimization.In order to solve this problem and acquire the particle's personal best position and the population's global best position in the multi-objective searching space,the non-dominated sorting algorithm is used to determine which result is better according to the multi-objective function values.A set named global best archive(GBA)is used to store all these non-dominated solutions and the Sigma method is utilized to find best local guides from the GBA for each particle of the population.The endmember extraction result is a Pareto front.(3)An endmember bundle extraction algorithm based on the iterative quantum-behaved particle swarm optimization is proposed.In recent years,although several endmember bundle extraction methods that considering endmember variability have been developed,most of them are still based on the convex geometry of HSIs.However,the convex geometry assumption is violated when there are noise and spectral variability in the image.Hence,the existing endmember bundle extraction methods have their limitations.In this thesis,we present a new framework called iterative endmember bundle extraction(IEBE),that determines endmember sets from the image using iterative implementations of QPSO method.Different from the conventional methods that identify endmembers based on either the shape of a spectral curve or the geometry of the data simplex,the proposed IEBE method extracts multiple endmember candidates by the iterative analysis of the root-mean-square error.The final endmember bundle set is obtained by combining the endmembers of each iteration.(4)The three proposed endmember extraction methods are compared,and the advantages and disadvantages of each method are summarized.Experimental results show that,the endmember bundle extraction method can achieve higher unmixing accuracy when there is spectral variability in the image;the endmember extraction method with high efficiency is more appropriate when the spectral variability can be ignored.