Scholarlink

<p>Hyperspectral image (HSI) classification remains a challenging task due to the high spectral dimensionality and the need for effective spatial feature integration. To address this, we propose a lightweight yet effective deep learning architecture named Patchwise Spectral-Spatial MambaNet (PatchMamba) that jointly models local spatial context and global spectral dependencies. The framework first extracts fixed-size local patches from the input hyperspectral cube and encodes spatial features using two-dimensional convolutional layers. These representations are reshaped into token sequences and passed through a stack of Spectral-Spatial Mamba (SS-Mamba) blocks, each composed of dense layers, layer normalization, and residual connections. A global average pooling layer aggregates the refined token features, and a final softmax classifier produces the predicted land-cover labels. Experimental testing on the Pavia University data demonstrates that PatchMamba attains a top accuracy of 99.08% and recalls of 98.11% and outperforms FC-DNN and non-patch-based SS-Mamba baselines in addition to generating better spatial coherence in classification maps. Both quantitative and qualitative results confirm the robustness and efficiency of the proposed method, making it a competitive solution for real-world HSI classification tasks.</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

A PatchMamba-based deep learning network for hyperspectral image classification

  • 期刊论文

  • 发表日期：

    2026年4月20日

查看全文
  • Mohammed Chachan Younis

摘要

Hyperspectral image (HSI) classification remains a challenging task due to the high spectral dimensionality and the need for effective spatial feature integration. To address this, we propose a lightweight yet effective deep learning architecture named Patchwise Spectral-Spatial MambaNet (PatchMamba) that jointly models local spatial context and global spectral dependencies. The framework first extracts fixed-size local patches from the input hyperspectral cube and encodes spatial features using two-dimensional convolutional layers. These representations are reshaped into token sequences and passed through a stack of Spectral-Spatial Mamba (SS-Mamba) blocks, each composed of dense layers, layer normalization, and residual connections. A global average pooling layer aggregates the refined token features, and a final softmax classifier produces the predicted land-cover labels. Experimental testing on the Pavia University data demonstrates that PatchMamba attains a top accuracy of 99.08% and recalls of 98.11% and outperforms FC-DNN and non-patch-based SS-Mamba baselines in addition to generating better spatial coherence in classification maps. Both quantitative and qualitative results confirm the robustness and efficiency of the proposed method, making it a competitive solution for real-world HSI classification tasks.