KL-Weighted Graph Sparse Self-Representation for Unsupervised Hyperspectral Band Selection

Unsupervised band selection (BS) methods have attracted much attention in hyperspectral imagery (HSI), which can select informative bands to solve the problems of information redundancy and high computational complexity. In this paper, we propose a KL-weighted graph sparse self-representation (SSR)...

Full description

Saved in:
Bibliographic Details
Published in2023 International Conference on Image Processing, Computer Vision and Machine Learning (ICICML) pp. 716 - 720
Main Authors Li, Pengjie, Zhao, Juan, Teng, Weike, Bai, Xia, Wang, Shaobo
Format Conference Proceeding
LanguageEnglish
Published IEEE 03.11.2023
Subjects
band selection (BS)
hyperspectral imagery (HIS)
Hyperspectral imaging
Kullack-Leibler (KL) divergence
Machine learning
Optimization
Redundancy
Sparse approximation
Sparse matrices
sparse self-representation (SSR)
Weight measurement
Online AccessGet full text
DOI10.1109/ICICML60161.2023.10424811

Cover

More Information
Summary:Unsupervised band selection (BS) methods have attracted much attention in hyperspectral imagery (HSI), which can select informative bands to solve the problems of information redundancy and high computational complexity. In this paper, we propose a KL-weighted graph sparse self-representation (SSR) method for unsupervised BS, in which the dissimilarity of bands measured via KL divergence is integrated into the superpixel-based graph SSR model by weighting the sparse representation coefficient matrix. An alternating optimization algorithm is designed to obtain the optimal coefficient matrix and the representative bands are finally selected by the norm ranking of rows of the coefficient matrix. Experimental results on HIS datasets show the effectiveness of the proposed BS algorithm and it outperforms other related BS methods in selecting proper representative bands for classification.
DOI:10.1109/ICICML60161.2023.10424811