Unsupervised Hyperspectral Unmixing Based on Constrained Nonnegative Matrix Factorization and Particle Swarm Optimization

The constrained nonnegative matrix factorization algorithm (CNMF) has previously been shown to be a useful method to solve the unmixing problem in hyper spectral remote sensing images, but it also has some key weaknesses which affect its applied range. It´s sensitive to the initial values, and easily falls to the local minimum. To solve the problems, a new intelligent optimization method - PSO(Particle Swarm Optimization) algorithm is combined with CNMF. The end members and abundance fractions obtained by CNMF are adopted as the initial values of PSO, the optimal solution of PSO is in reverse as the new initial value in the next running of CNMF, and this procedure is repeated until the global optimal solution is achieved. The experimental results based on synthetic data and real images demonstrate that the proposed method outperforms the standard CNMF algorithm and CNMF with output of VCA as its initial values.