EBM3GP: A novel evolutionary bi-objective genetic programming for dimensionality reduction in classification of hyperspectral data

Created by W.Langdon from gp-bibliography.bib Revision:1.7630

@Article{ZHOU:2023:infrared,

• author = "Zheng Zhou and Yu Yang and Gan Zhang and Libing Xu and Mingqing Wang",
• title = "{EBM3GP:} A novel evolutionary bi-objective genetic programming for dimensionality reduction in classification of hyperspectral data",
• journal = "Infrared Physics \& Technology",
• volume = "129",
• pages = "104577",
• year = "2023",
• ISSN = "1350-4495",
• DOI = "doi:10.1016/j.infrared.2023.104577",
• URL = "https://www.sciencedirect.com/science/article/pii/S135044952300035X",
• keywords = "genetic algorithms, genetic programming, Dimensionality reduction, Hyperspectral image, Low- and high-level features, Shannon entropy, Cross entropy",
• abstract = "Dimensionality reduction (DR) is vital in hyperspectral image (HSI) classification, and feature extraction and band selection methods have been demonstrated to be effective at accomplishing it. However, both types of methods can only obtain single-level features from HSI spectra, which suffers from insufficient useful information and makes accurate classification of the high dimensionality of HSI pixels challenging. To overcome the shortcomings, this study proposes a novel Evolutionary Bi-objective Genetic Programming-based unsupervised DR approach named EBM3GP for obtaining low-level features (bands) and high-level features from raw HSI spectra simultaneously. In EBM3GP, multi-dimensional trees are used to encode the raw spectrum to low- and high-level features; two mutually restrictive measures are applied to evaluate the amount of information and the redundancy contained in trees (evaluation does not use the HSI pixel's label); multiple trees are optimized through population evolution by combining three types of crossover operators, two types of mutation operators and nondominated sorting method; a Pareto optimal individual is finally output and decoded as a DR strategy. Then, this study applies Random Forest, least squares Support Vector Machine, and Extreme Learning Machine for classification to evaluate the efficacy of the DR strategy. Based on three HSIs (including Indian Pines, Salinas, and Pavia University datasets), EBM3GP is demonstrated to outperform five popular DR methods for HSI classification. Moreover, the EBM3GP is not sensitive to data size and thus is available for DR of small-size HSI datasets",

}

Genetic Programming entries for Zheng Zhou Yu Yang Gan Zhang Libing Xu Mingqing Wang

Citations