Scholarlink

<p>Remote sensing (RS) image classification plays a vital role in intelligent RS image interpretation. However, contemporary deep neural network-based methods heavily rely on large labeled datasets, which constrains performance improvement. To address this issue, this paper proposes a multi-head attention generative adversarial network (MAGAN) for RS image classification. First, considering the unique characteristics of RS images, we design a novel multi-head attention mechanism (MAM) to effectively model spatial relationships between pixels. Then, to cope with limited training data, we construct a novel generative adversarial network framework with both generator and discriminator equipped with MAM. The generator incorporates MAM for hierarchical feature synthesis, while the discriminator uses MAM to enhance spatial discrimination. The model is trained using original RS images through adversarial training until convergence. Finally, generated images from the multi-head attention generator are combined with original samples to train a robust classification model. Extensive experiments conducted on six public datasets, namely AID, RSI-CB, WHU-RS19, PatternNet, UCM, and NWPU-RESISC45, demonstrate that MAGAN generates high-quality synthetic images and significantly enhances classification performance of remote sensing images.</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

MAGAN: multi-head attention generative adversarial network for remote sensing image classification

  • 期刊论文

  • 发表日期：

    2025年5月12日

查看全文
  • Ke Li,
  • Zhonghua Luo,
  • Keyong Shen,
  • Jie Lu,
  • Zhijun Li

摘要

Remote sensing (RS) image classification plays a vital role in intelligent RS image interpretation. However, contemporary deep neural network-based methods heavily rely on large labeled datasets, which constrains performance improvement. To address this issue, this paper proposes a multi-head attention generative adversarial network (MAGAN) for RS image classification. First, considering the unique characteristics of RS images, we design a novel multi-head attention mechanism (MAM) to effectively model spatial relationships between pixels. Then, to cope with limited training data, we construct a novel generative adversarial network framework with both generator and discriminator equipped with MAM. The generator incorporates MAM for hierarchical feature synthesis, while the discriminator uses MAM to enhance spatial discrimination. The model is trained using original RS images through adversarial training until convergence. Finally, generated images from the multi-head attention generator are combined with original samples to train a robust classification model. Extensive experiments conducted on six public datasets, namely AID, RSI-CB, WHU-RS19, PatternNet, UCM, and NWPU-RESISC45, demonstrate that MAGAN generates high-quality synthetic images and significantly enhances classification performance of remote sensing images.