Dual-Branch Attention Convolution Spectral–Spatial Feature Extraction Networks for Hyperspectral Image Classification

The attention-based dual-branch network demonstrates significant potential in hyperspectral image (HSI) classification. However, the existing dual-branch approaches still face two crucial issues that hinder the further performance promotion of HSI classification: 1) The static convolution in feature...

Full description

Saved in:
Bibliographic Details
Main Authors: Genyun Sun, Cheng Jing, Aizhu Zhang, Xiaolin Chen, Hang Fu, Zhaojie Pan, Ji Cheng
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Subjects:
Online Access:https://ieeexplore.ieee.org/document/11048514/
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:The attention-based dual-branch network demonstrates significant potential in hyperspectral image (HSI) classification. However, the existing dual-branch approaches still face two crucial issues that hinder the further performance promotion of HSI classification: 1) The static convolution in feature extraction branches is incapable of dynamically exploring the local and global dependency features in HSI. 2) Dual-branch networks tend to overlook the fusion of deep spectral–spatial semantic information. To address these challenges, a dual-branch attention convolution spectral–spatial feature extraction network (DACSS) is proposed. DACSS encompasses a dual-branch feature extraction module (DFEM) and a spatial–spectral feature aggregation module (SSAM). DFEM dynamically extracts different scales features from both spatial and spectral branches by combining attention convolution and spectral–spatial attention mechanisms. Within these components, the attention convolution module adaptively highlights the extraction of detailed regional features in HSI, while the spectral–spatial attention module effectively alleviates the limitations of long-range spatial dependencies arising from the convolution process. Based on the attention weights, SSAM performs multiscale fusion of deep semantic information from both spatial and spectral branches. The experimental results show that the proposed method outperforms several state-of-the-art methods, achieving overall accuracies of 91.69%, 89.24%, 89.50%, and 94.67% on the HongHu, SG, UH2018, and QUH-TDW datasets, respectively.
ISSN:1939-1404
2151-1535