Sparse Unmixing for Hyperspectral Imagery via Comprehensive-Learning-Based Particle Swarm Optimization

Sparse unmixing methods have been extensively studied as a popular topic in hyperspectral image analysis for several years. Fundamental model-based unmixing problems can be better reformulated by exploiting sparse constraints in different forms. Gradient-based optimization approaches commonly serve for traditional sparse unmixing, but their limitations such as one-way search, often induce unsatisfactory local optimum, especially when the problems are nonconvex. Therefore, acceptable unmixing performance cannot always be guaranteed, and the sparsity of hyperspectral imagery may be incorrectly expressed. In this article, an unsupervised sparse unmixing method using comprehensive-learning-based particle swarm optimization (PSO) is proposed. Due to the basic PSO's premature convergence in dealing with high-dimensional problems, double swarms whose fitness functions are accordingly divided into a series of low-dimensional subproblems are constructed to search for optimal endmembers and abundances alternately, leading to the implementation of unmixing in refined solution spaces. Under this framework, two comprehensive learning strategies are introduced to promote and refine particles’ mutual learning deeply at the element-level, through which the abundance sparsity in every local pixel and every endmember's global abundance sparsity can be better exploited and expressed. Experiments with both simulated datasets and real hyperspectral images are employed to validate the performance of the proposed method combined with different sparse constraints. In comparison with other state-of-the-art algorithms, the proposed method enables the achievement of better unmixing results.