Scholarlink

<p>As one of the most powerful techniques in hyperspectral image (HSI) classification, sparse subspace clustering (SSC) utilizes the self-representation principle to uncover the underlying low-rank structures of high-dimensional spectral signatures. However, the efficacy of traditional SSC is frequently compromised by the lack of physical semantic guidance and the rigid nature of first-order spatial modeling. In this article, a target-aware locally guided sparse subspace clustering paradigm with high-order spatial constraints is established for unsupervised hyperspectral image classification. First, a target-aware unmixing mechanism is formulated to perceive latent physical attributes and a target-guided matrix is then constructed to steer the sparse representation toward localized manifolds with high physical fidelity, effectively mitigating the interference of false positive neighbors. Second, a high-order spatial regularization term derived from superpixel segmentation is integrated to enforce structural consistency and preserve non-rigid geometric boundaries in complex scenes. Third, a synergistic joint optimization framework is developed to establish a reciprocal feedback loop between signal purification and attribute perception, where the reconstructed purified signal drives the refinement of semantic priors. Finally, experimental results have validated the performance of the proposed work on the Indian Pines and Pavia University datasets.</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

Target-aware locally guidance matrix for unsupervised hyperspectral image classification

  • 期刊论文

  • 发表日期：

    2026年6月26日

查看全文
  • Xia Wei,
  • Zhiwei Li

摘要

As one of the most powerful techniques in hyperspectral image (HSI) classification, sparse subspace clustering (SSC) utilizes the self-representation principle to uncover the underlying low-rank structures of high-dimensional spectral signatures. However, the efficacy of traditional SSC is frequently compromised by the lack of physical semantic guidance and the rigid nature of first-order spatial modeling. In this article, a target-aware locally guided sparse subspace clustering paradigm with high-order spatial constraints is established for unsupervised hyperspectral image classification. First, a target-aware unmixing mechanism is formulated to perceive latent physical attributes and a target-guided matrix is then constructed to steer the sparse representation toward localized manifolds with high physical fidelity, effectively mitigating the interference of false positive neighbors. Second, a high-order spatial regularization term derived from superpixel segmentation is integrated to enforce structural consistency and preserve non-rigid geometric boundaries in complex scenes. Third, a synergistic joint optimization framework is developed to establish a reciprocal feedback loop between signal purification and attribute perception, where the reconstructed purified signal drives the refinement of semantic priors. Finally, experimental results have validated the performance of the proposed work on the Indian Pines and Pavia University datasets.