Research On Multi-Objective Particle Swarm Optimization Based Nonlinear Unmixing Methods For Hyperspectral Remote Sensing Images

The availability of hyperspectral remote sensing technology has made continuous spectral imaging of materials in tens to hundreds of continuous bands a possibility.The hyperspectral remote sensing images acquired by this technology carry rich spatial and spectral information,which effectively promotes fine diagnosis and identification of features.However,the phenomenon of mixed pixels is also particularly prominent in hyperspectral images when a large number of redundant bands puts significant pressure on the efficient processing of data.The limited spatial resolution makes many different materials often exist in a single observation unit,and the spectral features of the mixed pixels formed by the joint action of related features are captured by the sensor.Therefore,decomposition of pixels of hyperspectral images into endmembers and abundances using fine feature information extracted by spectral unmixing techniques is very important for improving practical remote sensing applications.To account for the nonlinear mixing effects caused by multiple scattering of light in real complex scenes,the numerical optimization problem of nonlinear unmixing is more difficult and tricky to solve than the traditional simple linear unmixing.The excellent performance of intelligent algorithms such as particle swarm optimization in solving problems including complicated constrained problems and nonconvex nonlinear problems has enabled it to be a feasible means to solve nonlinear spectral unmixing problems.In addition,multi-objective optimization techniques often provide better trade-offs than the traditional fixed penalty coefficient method and can be easily combined with particle swarm optimization to guide the correct search of the algorithm for various normal terms such as sparse parametrization,which are often introduced in spectral unmixing problems.Therefore,this paper intends to use the framework of particle swarm combined with multi-objective optimization for unmixing as the basis to reduce the effect of colinearity effects in nonlinear mixture models and to enhance the task of expressing abundance sparsity in nonlinear unmixing in two aspects.The corresponding effective improvement strategies and mechanisms are also designed to enhance the unmixing accuracy.The main research of this work is as follows:1.A nonlinear unmixing method based on multi-objective particle swarm and geometric projection improvement is proposed.Firstly,the geometric projection is used in combination with a multi-objective particle swarm framework to mitigate the effects of colinearity effects among endmembers and to improve the constraint handling of unsupervised nonlinear unmixing based on polynomial posterior nonlinear mixture models.Secondly,the multi-objective particle swarm algorithm framework is improved by a periodic reset mechanism,which not only can dynamically enhance its connection with geometric projection but also can jump out of the local optimum by using the property that the particle swarm algorithm is not easily trapped in the local optimum,while the information exchange among particles in the swarm can be well guided to ensure the convergence and search accuracy of the unmixing algorithm.By conducting comparison experiments on simulated data and real hyperspectral remote sensing image data,the experimental results show that the proposed algorithm in this paper outperforms the compared algorithm and verifies the effectiveness of the algorithm.2.A nonlinear sparse unmixing method based on multi-objective particle swarm optimization is proposed.At first,to further increase the applicability of nonlinear unmixing to complex scenes effectively,a multilinear mixture model is used as the main body to construct the nonlinear unmixing problem.Then,a bi-objective term on abundance optimization is constructed based on the sparsity of the hyperspectral image’s pixel abundances,and a novel multi-objective optimization strategy based on dominance ranking is used to balance the feasibility and optimality of particles to guide the search of the population by weighting.In addition,the reconstruction error is directly used as the fitness function to guide another swarm’s search endmembers.By conducting experiments on both simulated data with sparsity and real hyperspectral remote sensing image data,the proposed algorithms in this paper obtain the best unmixing results and demonstrate that the multi-objective particle swarm algorithm combining abundance sparsity can further improve the accuracy and good applicability of unmixing.