Scholarlink

<p>Hyperspectral image (HSI) classification is a key task in the field of remote sensing with significant practical value. However, most convolutional neural network (CNN)-based methods primarily focus on capturing local spatial features while overlooking global context. In contrast, vision transformer (ViT)-based methods can effectively model long-range dependencies but fail to adequately capture detailed local spatial structures and lack explicit spatial priors in their self-attention mechanism. In this paper, we propose a novel CNN–ViT hybrid architecture for HSI classification, called the Manhattan self-attention transformer with adaptive global–local channel attention network (MTACANet). This network can simultaneously extract local and global features and incorporates an explicit, distance-based spatial prior into the attention mechanism. First, we introduce a wavelet-guided multi-scale spatial-spectral feature extraction (WMSFE) block to separately extract multi-scale spatial and spectral features, which begins with wavelet transform convolution (WTConv) featuring an expanded receptive field to strengthen feature representation capabilities. Second, we use a Manhattan self-attention transformer encoder (MSATE) to model long-range spatial dependencies across the extracted multi-scale spatial feature representations. The MSATE incorporates explicit spatial priors and adjusts attention weights based on inter-token distances, enabling each target token to focus more on nearby tokens while suppressing attention to distant ones. In addition, an adaptive global-local channel attention (AGLCA) module is employed to facilitate interactions between local and global channel information, allowing for more precise extraction of fine-grained spectral characteristics in HSIs. Experiments on the Indian Pines, GF-5 Yancheng, ZY1-02D Huanghekou, and WHU-Hi-HongHu datasets show that its overall accuracy reaches 99.11%, 99.28%, 99.09%, and 95.57%, respectively, exhibiting excellent performance. Our code is available at: <a href="https://github.com/zhe-meng/MTACANet">https://github.com/zhe-meng/MTACANet</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

Hyperspectral image classification via Manhattan self-attention transformer and adaptive global-local channel attention

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月28日

查看全文
  • Zhe Meng,
  • Pan Yue,
  • Feng Zhao

摘要

Hyperspectral image (HSI) classification is a key task in the field of remote sensing with significant practical value. However, most convolutional neural network (CNN)-based methods primarily focus on capturing local spatial features while overlooking global context. In contrast, vision transformer (ViT)-based methods can effectively model long-range dependencies but fail to adequately capture detailed local spatial structures and lack explicit spatial priors in their self-attention mechanism. In this paper, we propose a novel CNN–ViT hybrid architecture for HSI classification, called the Manhattan self-attention transformer with adaptive global–local channel attention network (MTACANet). This network can simultaneously extract local and global features and incorporates an explicit, distance-based spatial prior into the attention mechanism. First, we introduce a wavelet-guided multi-scale spatial-spectral feature extraction (WMSFE) block to separately extract multi-scale spatial and spectral features, which begins with wavelet transform convolution (WTConv) featuring an expanded receptive field to strengthen feature representation capabilities. Second, we use a Manhattan self-attention transformer encoder (MSATE) to model long-range spatial dependencies across the extracted multi-scale spatial feature representations. The MSATE incorporates explicit spatial priors and adjusts attention weights based on inter-token distances, enabling each target token to focus more on nearby tokens while suppressing attention to distant ones. In addition, an adaptive global-local channel attention (AGLCA) module is employed to facilitate interactions between local and global channel information, allowing for more precise extraction of fine-grained spectral characteristics in HSIs. Experiments on the Indian Pines, GF-5 Yancheng, ZY1-02D Huanghekou, and WHU-Hi-HongHu datasets show that its overall accuracy reaches 99.11%, 99.28%, 99.09%, and 95.57%, respectively, exhibiting excellent performance. Our code is available at: https://github.com/zhe-meng/MTACANet.