Dimensionality Reduction Using Band Optimisation

Dimensionality reduction of hyperspectral imagery is an important pre-processing step for achieving improved classification accuracy. In this chapter, optimisation-based band selection approaches are discussed using genetic algorithms (GAs) and particle swam optimisation (PSO). In contrast to exhaustive search algorithms, optimisation-based approach employs fast search measures to find a better solution in a large solution space. The key strength of the optimisation-based technique is that it evaluates the quality of features in a group instead of considering the merits of individual features. Thus, redundancy among the selected bands can be avoided. The divergence/information theory–based objective function along with the image spatial information are used for the fitness evaluation of features in the unsupervised band optimisation techniques. This yielded comparably better classification accuracies and improved the speed of convergence. The reduction of spatial dimension also contributed in improving the computation time. The presented methods also reduced the effect of noise in hyperspectral imagery for efficient band selection.