Scholarlink

<p>Hyperspectral image classification across different domains remains a challenging task due to data distribution shifts. Traditional methods often struggle with domain adaptation and generalization. To address these issues, we propose a novel multisource domain generalization method based on joint-product distribution alignment and supervised contrastive learning (JPDA-SCL). JPDA-SCL leverages joint-product distribution alignment (JPDA) to eliminate domain labels and achieve global domain alignment, while supervised contrastive learning (SCL) enhances feature distinguishability and prevents class feature collapse. Specially, relative Chi-square divergence measures distribution differences to simplify domain alignment. In addition, a two-branch network, incorporating attribute extraction and integrated feature extraction modules, captures effective multilayer features of hyperspectral images. Experimental results on four public datasets demonstrate the superiority of JPDA-SCL, achieving improvements of 1% to 5% in overall accuracy, 1–5% in average accuracy, and 1–6% in Kappa coefficient compared to competitive methods. Here, we show that JPDA-SCL effectively mitigates spectral domain shift issues, providing a robust solution for hyperspectral image classification in diverse domains. The code will be made available at: <a href="https://github.com/LLL2008/JPDA-SCL.">https://github.com/LLL2008/JPDA-SCL.</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

Enhancing domain generalization in hyperspectral image classification via joint-product distribution alignment and supervised contrastive learning

  • 期刊论文

  • 发表日期：

    2025年12月13日

查看全文
  • Gaojian Luo,
  • Xiaolan Liu,
  • Mengying Xie,
  • Pei Huang

摘要

Hyperspectral image classification across different domains remains a challenging task due to data distribution shifts. Traditional methods often struggle with domain adaptation and generalization. To address these issues, we propose a novel multisource domain generalization method based on joint-product distribution alignment and supervised contrastive learning (JPDA-SCL). JPDA-SCL leverages joint-product distribution alignment (JPDA) to eliminate domain labels and achieve global domain alignment, while supervised contrastive learning (SCL) enhances feature distinguishability and prevents class feature collapse. Specially, relative Chi-square divergence measures distribution differences to simplify domain alignment. In addition, a two-branch network, incorporating attribute extraction and integrated feature extraction modules, captures effective multilayer features of hyperspectral images. Experimental results on four public datasets demonstrate the superiority of JPDA-SCL, achieving improvements of 1% to 5% in overall accuracy, 1–5% in average accuracy, and 1–6% in Kappa coefficient compared to competitive methods. Here, we show that JPDA-SCL effectively mitigates spectral domain shift issues, providing a robust solution for hyperspectral image classification in diverse domains. The code will be made available at: https://github.com/LLL2008/JPDA-SCL.