Scholarlink

<p>Classifying hyperspectral images (HSIs) is a key challenge in remote sensing, with convolutional neural networks (CNNs) and transformer models becoming leading techniques in this area. CNNs, while effective, often struggle to adequately capture intricate semantic features, and increasing network depth leads to significantly higher computational costs. Conversely, transformers, despite their efficacy in modeling spectral-spatial dependencies, introduce significant computational overhead due to their complexity. Mamba, leveraging the state space model (SSM), presents a compelling alternative that efficiently captures long-range dependencies in HSIs while ensuring computational efficiency with linear complexity. To improve the classification performance of HSIs by simultaneously extracting rich local and global spatial-spectral features, as well as deep semantic features, while reducing the computational complexity of the model, this paper proposes an innovative hybrid large selective kernel and convolutional additive self-attention model (HLSK-CASMamba) for HSI classification. First, we design a feature extraction module that combines a 3D convolution layer, a 2D convolution layer, and a large selective kernel (LSK) network, enabling the efficient extraction of both depth-related and spatial details information from HSIs. Second, we propose a novel CASMamba model, with its core module, CAS-VSSM, combining convolutional additive self-attention (CAS) and the vision state-space sequence model (VSSM). This fusion leverages the local feature extraction of convolutions, spatial dependency modeling of self-attention, and long-range dependency handling of VSSM, enhancing the capture of both local and global context while ensuring computational efficiency. Finally, we incorporate the KANLinear module to replace the traditional linear layer, enhancing sample label acquisition. Extensive evaluations on three popular HSIs show that, under 10% training samples, the proposed method achieves 99.57% accuracy on the Houston 2013 dataset, 99.96% on the Botswana dataset, and 99.92% on the University of Pavia dataset, outperforming various existing advanced techniques.</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

HLSK-CASMamba: hybrid large selective kernel and convolutional additive self-attention mamba for hyperspectral image classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年6月2日

查看全文
  • Xiaoqing Wan,
  • Yupeng He,
  • Feng Chen,
  • Ziqi Sun,
  • Dongtao Mo

摘要

Classifying hyperspectral images (HSIs) is a key challenge in remote sensing, with convolutional neural networks (CNNs) and transformer models becoming leading techniques in this area. CNNs, while effective, often struggle to adequately capture intricate semantic features, and increasing network depth leads to significantly higher computational costs. Conversely, transformers, despite their efficacy in modeling spectral-spatial dependencies, introduce significant computational overhead due to their complexity. Mamba, leveraging the state space model (SSM), presents a compelling alternative that efficiently captures long-range dependencies in HSIs while ensuring computational efficiency with linear complexity. To improve the classification performance of HSIs by simultaneously extracting rich local and global spatial-spectral features, as well as deep semantic features, while reducing the computational complexity of the model, this paper proposes an innovative hybrid large selective kernel and convolutional additive self-attention model (HLSK-CASMamba) for HSI classification. First, we design a feature extraction module that combines a 3D convolution layer, a 2D convolution layer, and a large selective kernel (LSK) network, enabling the efficient extraction of both depth-related and spatial details information from HSIs. Second, we propose a novel CASMamba model, with its core module, CAS-VSSM, combining convolutional additive self-attention (CAS) and the vision state-space sequence model (VSSM). This fusion leverages the local feature extraction of convolutions, spatial dependency modeling of self-attention, and long-range dependency handling of VSSM, enhancing the capture of both local and global context while ensuring computational efficiency. Finally, we incorporate the KANLinear module to replace the traditional linear layer, enhancing sample label acquisition. Extensive evaluations on three popular HSIs show that, under 10% training samples, the proposed method achieves 99.57% accuracy on the Houston 2013 dataset, 99.96% on the Botswana dataset, and 99.92% on the University of Pavia dataset, outperforming various existing advanced techniques.