.png){kind=icon}
.png){kind=icon}
Return
Hyperspectral Band Selection via Heterogeneous Graph Convolutional Self-Representation Network
.png){kind=icon}
OA
AI
DOI：10.1109/JSTARS.2025.3589866.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Hyperspectral image (HSI) band selection (BS) plays a crucial role in HSI dimensionality reduction, aiming to identify a representative subset of bands with minimal redundancy. However, conventional BS approaches primarily operate in the Euclidean domain, often overlooking the structural characteristics of pixels and spectral bands, such as spatial continuity and spectral dependencies. In addition, they handle each HSI as an integrated unit to harness implicit spatial information, disregarding spatial distribution variations across different homogeneous regions. To fully leverage structural information, this study introduces a novel BS method, termed the dual heterogeneous graph convolutional network with enhanced self-representation (ESR-HGCN), for HSI BS. The heterogeneous graph convolutional network (HGCN) and enhanced self-representation (ESR) serve as the two essential components of the proposed ESR-HGCN. To explore spatial features and the potential hidden interactions among spectral bands, we use the HGCN as the backbone network for heterogeneous graph-based HSI BS. Dual graphs at the pixel and band levels are separately constructed and integrated into the ESR module, where a sparsity constraint is enforced and the original Frobenius norm is replaced with <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\ell _{1}$</tex-math></inline-formula>- and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\ell _{2,1}$</tex-math></inline-formula>-norm regularizations to achieve robust BS. Meanwhile, dual graph convolution operations are performed to separately extract spatial and spectral features, thereby seamlessly integrating spectral, spatial, and geometric information, offering significant advantages for HSI BS. Finally, an effective optimization scheme is developed to refine the proposed approach. Experimental findings on representative HSI datasets highlight the superiority of ESR-HGCN over state-of-the-art methods.
Keywords:
Band selection (BS)
graph convolutional network (GCN)
hyperspectral image (HSI)
self-representation (SR)
Journal
IF:
5.3
Papers:
1.3K
Citations:
3.0W