Scholarlink

<p>In recent years, Transformers have emerged as powerful tools for hyperspectral image (HSI) classification due to their ability to capture vglobal features. Howeer, Transformers tend to overlook local features, and most existing methods rely heavily on labeled data. To address these limitations, this paper proposes a Masked Self-supervised Dual-scale Transformer (MSDT) for HSI classification. MSDT introduces a dual-scale encoder network that extracts features at different scales, enabling the model to capture both global and local features. Additionally, a class token exchange operation facilitates the fusion of features across scales. A masked self-supervised pretraining approach is employed to learn rich feature representations from unlabeled HSI data, further enhancing classification performance. Experimental results on several benchmark datasets demonstrate that MSDT achieves state-of-the-art classification accuracy, outperforming baseline methods. The proposed method not only improves the utilization of unlabeled data but also provides a new perspective for HSI classification using Transformers. The full source code of the proposed MSDT can be found at “<a href="https://github.com/MSDT-main/MSDT">https://github.com/MSDT-main/MSDT</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

Masked self-supervised dual-scale transformer for enhanced hyperspectral image classification

  • 期刊论文

  • 发表日期：

    2025年9月3日

查看全文
  • Caihong Mu,
  • Zeyu Zhang,
  • Yafeng Wang,
  • Suling Chen,
  • Yi Liu

摘要

In recent years, Transformers have emerged as powerful tools for hyperspectral image (HSI) classification due to their ability to capture vglobal features. Howeer, Transformers tend to overlook local features, and most existing methods rely heavily on labeled data. To address these limitations, this paper proposes a Masked Self-supervised Dual-scale Transformer (MSDT) for HSI classification. MSDT introduces a dual-scale encoder network that extracts features at different scales, enabling the model to capture both global and local features. Additionally, a class token exchange operation facilitates the fusion of features across scales. A masked self-supervised pretraining approach is employed to learn rich feature representations from unlabeled HSI data, further enhancing classification performance. Experimental results on several benchmark datasets demonstrate that MSDT achieves state-of-the-art classification accuracy, outperforming baseline methods. The proposed method not only improves the utilization of unlabeled data but also provides a new perspective for HSI classification using Transformers. The full source code of the proposed MSDT can be found at “https://github.com/MSDT-main/MSDT”.