Scholarlink

<p>Model-agnostic meta-learning has garnered significant attention in remote sensing (RS) scene classification due to its flexibility and generalization capability. However, RS images are often influenced by varying imaging platforms, resolutions, and acquisition conditions, leading to substantial discrepancies across tasks. These variations hinder the generalization performance of conventional meta-learning approaches. To address this issue, we propose a Variational Meta-Learning (VML) framework based on variational inference. VML generates task-adaptive model parameters by leveraging latent task features, thereby improving the model’s adaptability across diverse task environments. Specifically, we introduce the concept of a task central point, which serves as a latent distribution representation of meta-parameters. This task-specific distribution is then used to guide model initialization, enhancing the capacity to model task variability in RS scenarios. To evaluate the effectiveness of the proposed method, we conduct extensive experiments on four publicly available RS datasets: UC Merced, WHU-RS19, NWPU-RESISC45, and EuroSAT. Results demonstrate that VML consistently outperforms representative meta-learning baselines under various few-shot settings, particularly in scenarios with high task distribution diversity.</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

Few-shot remote sensing image scene classification based on variational meta-learning

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Jin Wang,
  • Xunhang Shang,
  • Bo Yang,
  • Haipeng Ren,
  • Hong Jiang,
  • Hongge Yao,
  • Wei Li

摘要

Model-agnostic meta-learning has garnered significant attention in remote sensing (RS) scene classification due to its flexibility and generalization capability. However, RS images are often influenced by varying imaging platforms, resolutions, and acquisition conditions, leading to substantial discrepancies across tasks. These variations hinder the generalization performance of conventional meta-learning approaches. To address this issue, we propose a Variational Meta-Learning (VML) framework based on variational inference. VML generates task-adaptive model parameters by leveraging latent task features, thereby improving the model’s adaptability across diverse task environments. Specifically, we introduce the concept of a task central point, which serves as a latent distribution representation of meta-parameters. This task-specific distribution is then used to guide model initialization, enhancing the capacity to model task variability in RS scenarios. To evaluate the effectiveness of the proposed method, we conduct extensive experiments on four publicly available RS datasets: UC Merced, WHU-RS19, NWPU-RESISC45, and EuroSAT. Results demonstrate that VML consistently outperforms representative meta-learning baselines under various few-shot settings, particularly in scenarios with high task distribution diversity.