Evolutionary Multitasking Optimization for Multiobjective Hyperspectral Band Selection

Abstract

Hyperspectral remote sensing technology combines imaging technology and spectral technology, which greatly promotes the development of remote sensing science. However, the large amount of data, high redundancy and high data dimension of hyperspectral images will cause many problems such as “curse of dimensionality”. Feature extraction and band selection are two main methods to reduce the dimensionality and retain information for practical application. Compared with the feature extraction, band selection aims to select a band subset to reduce dimensionality, which can maintain the physical meaning of the original band. However, many band selection methods usually face many problems such as high computational cost, local optimization, poor classification accuracy and so on. In this paper, considering the evolutionary multitasking optimization algorithm has the characteristics of processing multiple tasks at the same time to improve the search efficiency, band selection is modeled as a multitasking optimization problem, and the evolutionary multitasking optimization algorithm is used as search strategy to select the band subset. Using hyperspectral remote sensing dataset as the experiment, the result shows that compared with other methods, the band subsets obtained by the evolutionary multitasking optimization algorithm have excellent overall classification accuracy and average classification accuracy, which will contribute to practical application.