A Diversified Multi-objective Particle Swarm Optimization Algorithm for Unsupervised Band Selection of Hyperspectral Images

Evolutionary algorithm (EA) with good search capability has been successfully extended to a mainstream band selection (BS) technique for hyperspectral images (HSIs). However, most of the existing methods still face two challenges: 1) falling into local optimum due to the single search strategy; 2) ignoring the problem of potential duplicate bands. To address these issues, this paper proposes an effective unsupervised BS method by using a diversified multi-objective particle swarm optimization (PSO) algorithm, called DPSO-BS. First, a new unsupervised BS model is designed, which applies the information entropy and structural similarity measure as two optimization objectives. Then, two complementary PSO search strategies are proposed to solve the above constructed BS model. In addition, a self-repair mechanism is designed to correct the offending solutions with duplicate bands. Experimental results on three HSI datasets demonstrate that DPSO-BS outperforms several state-of-the-art BS methods.