Scholarlink

<p>In hyperspectral image (HSI) classification, deep learning techniques have significantly advanced performance, yet challenges such as inadequate spatial–spectral feature extraction, high computational complexity, and inefficient multi-scale feature integration persist. To address these issues, we propose CFINet, a novel channel-enhanced feature integration network comprising three key modules: the spatial dual attention module (SDAM), the channel-enhanced cross-attention module (CECAM), and the multi-scale feature integration module (MSFIM). SDAM enhances spatial information interactions, CECAM captures inter-channel relationships, and MSFIM integrates features across multiple scales. Extensive experiments on four benchmark datasets demonstrate that CFINet significantly outperforms several state-of-the-art methods, achieving overall accuracies of 99.91%, 99.97%, 99.91%, and 99.90% on the Pavia University, Botswana, WHU-Hi-HanChuan, and WHU-Hi-HongHu datasets, respectively. This study highlights the effectiveness of hierarchical attention-driven feature integration in advancing HSI classification. The source codes are available at: <a href="https://github.com/wanxiaoqing/CFINet">https://github.com/wanxiaoqing/CFINet</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hierarchical attention-driven feature integration network for enhanced hyperspectral image classification

  • 期刊论文

  • 发表日期：

    2026年1月3日

查看全文
  • Zhize Li,
  • Xiaoqing Wan,
  • Feng Chen,
  • Hui Liu,
  • Dongtao Mo

摘要

In hyperspectral image (HSI) classification, deep learning techniques have significantly advanced performance, yet challenges such as inadequate spatial–spectral feature extraction, high computational complexity, and inefficient multi-scale feature integration persist. To address these issues, we propose CFINet, a novel channel-enhanced feature integration network comprising three key modules: the spatial dual attention module (SDAM), the channel-enhanced cross-attention module (CECAM), and the multi-scale feature integration module (MSFIM). SDAM enhances spatial information interactions, CECAM captures inter-channel relationships, and MSFIM integrates features across multiple scales. Extensive experiments on four benchmark datasets demonstrate that CFINet significantly outperforms several state-of-the-art methods, achieving overall accuracies of 99.91%, 99.97%, 99.91%, and 99.90% on the Pavia University, Botswana, WHU-Hi-HanChuan, and WHU-Hi-HongHu datasets, respectively. This study highlights the effectiveness of hierarchical attention-driven feature integration in advancing HSI classification. The source codes are available at: https://github.com/wanxiaoqing/CFINet.