An improved cuckoo search-based adaptive band selection for hyperspectral image classification

The information in hyperspectral images usually has a strong correlation, a large number of bands, which lead to the "curse of dimensionality". So, band selection is usually used to address this issue. However, problems remain for band selection, such as how to search for the most informative bands, and how many bands should be selected. In this paper, a cuckoo search (CS)-based adaptive band selection framework is proposed to simultaneously select bands and determine the optimal number of bands to be selected. The proposed framework includes two "cuckoo search", i.e. the outer one for estimating the optimal number of bands and the inner one for the corresponding band selection. To avoid employing an actual classifier within CS so as to greatly reduce computational cost, minimum estimated abundance covariance (MEAC) and Jeffreys-Matusita (JM) distance are adopted as criterion functions, which measures class separability. For the experiments, two widely used hyperspectral images, which acquired by the Hyperspectral digital imagery collection experiment (HYDICE) and the airborne Hyperspectral Mapper (HYMAP) system, are adopted for performance evaluation. The experimental results show that the two-CS-based algorithm outperforms the popular sequential forward selection (SFS), sequential floating forward search (SFFS), and other similar algorithms for hyperspectral band selection.