Scholarlink

<p>Hyperspectral image classification (HSIC) is an important task in remote sensing, enabling precise identification of materials across diverse applications from agriculture to urban planning. The intricate spatial–spectral structure of hyperspectral data presents significant challenges, including modeling long-range dependencies and capturing fine-grained details. While recent Mamba-based models show promise in sequence modeling, they often fall short in better exploiting spatial structures and inter-band correlations. Here, we propose MLFMamba, a novel multi-level feature modeling framework built upon the Mamba architecture. Our method progressively transitions from shallow local detail extraction to deep global semantic modeling. A Feature Aware (FA) module first captures multi-scale spatial features. Subsequently, a deep encoder employs two novel branches: a Spatial Multi-Scan Mamba (SpaMSM) module for rich spatial context capture via multi-directional scanning and a Spectral Grouping Bidirectional Mamba (SpeGBM) module for efficient modeling of spectral correlations. These are adaptively fused via a Spatial–Spectral Adaptive Fusion (SSAF) module. Comparative experiments on four benchmark datasets (Indian Pines, Pavia University, WHU-Hi-LongKou, and Houston 2013) demonstrate that MLFMamba achieves competitive performance. It improves overall accuracy by 0.87 to 1.59 percentage points over the well-performing competing methods while maintaining reasonable computational efficiency. This work underscores the potential of advanced state space models for complex spatial–spectral reasoning in remote sensing imagery. Our code has been open-sourced and is publicly available at <a href="https://github.com/foopy113/MLF">https://github.com/foopy113/MLF</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

Mamba unleashed: a multi-level feature modeling framework for enhanced hyperspectral image classification

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • Guicai Wang,
  • Fangfang Wang,
  • Xuan Fei,
  • Meng Wang

摘要

Hyperspectral image classification (HSIC) is an important task in remote sensing, enabling precise identification of materials across diverse applications from agriculture to urban planning. The intricate spatial–spectral structure of hyperspectral data presents significant challenges, including modeling long-range dependencies and capturing fine-grained details. While recent Mamba-based models show promise in sequence modeling, they often fall short in better exploiting spatial structures and inter-band correlations. Here, we propose MLFMamba, a novel multi-level feature modeling framework built upon the Mamba architecture. Our method progressively transitions from shallow local detail extraction to deep global semantic modeling. A Feature Aware (FA) module first captures multi-scale spatial features. Subsequently, a deep encoder employs two novel branches: a Spatial Multi-Scan Mamba (SpaMSM) module for rich spatial context capture via multi-directional scanning and a Spectral Grouping Bidirectional Mamba (SpeGBM) module for efficient modeling of spectral correlations. These are adaptively fused via a Spatial–Spectral Adaptive Fusion (SSAF) module. Comparative experiments on four benchmark datasets (Indian Pines, Pavia University, WHU-Hi-LongKou, and Houston 2013) demonstrate that MLFMamba achieves competitive performance. It improves overall accuracy by 0.87 to 1.59 percentage points over the well-performing competing methods while maintaining reasonable computational efficiency. This work underscores the potential of advanced state space models for complex spatial–spectral reasoning in remote sensing imagery. Our code has been open-sourced and is publicly available at https://github.com/foopy113/MLF.