Particle swarm optimization for nonlinear spectral unmixing: A case study of generalized bilinear model

Spectral unmixing is a key procedure of hyperspectral remote sensing analysis. Recently, nonlinear unmixing becomes a hotspot in this research field. However, due to the complexity of the nonlinear mixing models (nonLMM), it is trivial to develop an algorithm for nonlinear unmixing. Particle Swarm Optimization (PSO) is a classical algorithm of natural computation, which presents a great potential of nonlinear unmixing. Specially, it employs only a fitness value directly determined by the nonLMM, and does not need any information concerning about the gradient, hessian matrix, or probability distributions. Thus, it can be easily applied to characterize complex nonLMMs. In this paper, we develop a biswarm (double swarm) PSO algorithm for nonlinear unmixing, with a case study of generalized bilinear model (GBM). The experimental results indicate that the proposed algorithm outperforms other traditional algorithms for hyperspectral images. As a result, we can conclude that the PSO algorithm is an excellent method for addressing nonlinear unmixing problem.