Scholarlink

<p>Hyperspectral images (HSIs), which comprise numerous redundant spectral bands, are the most prevalent remote sensing sources for interpreting objects based on spectral band data. In the classification process, choosing a subset of spectral bands for data dimensionality reduction is known as band selection (BS). Deep learning (DL) based model with an attention mechanism can be used for BS, considering the nonlinear and global interaction among spectral bands. However, the existing DL based BS approaches using attention mechanisms are either unable to record both the spectral and spatial long-range information or depend only on queries, but not on keys and values. Moreover, the used reconstruction network (RecNet) in most of the studied BS techniques cannot recognize the features of images in compound scales since the network uses single-size kernels in convolution operations. To reduce these limitations of existing DL based BS approaches with attention mechanism, a novel DL based Hyperspectral BS (<Emphasis FontCategory="NonProportional">PSAA-MSRecNet</Emphasis>) model consisting of position-sensitive axial attention (PSAA) with a multi-scale RecNet (MSRecNet) has been proposed. It takes advantage of a PSAA module that adds not only queries but also keys and values-dependent positional bias terms. The MSRecNet is employed after the PSAA module as a RecNet to search the image features in divergent scales. The proposed approach is able to effectively capture quality feature representations and, as a consequence, pick the most informative bands with higher accuracy in classification tasks than prior studied BS approaches, the majority of cases accross three standard HSI 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

Deep learning based hyperspectral band selection using position-sensitive axial attention mechanism

  • 期刊论文

  • 发表日期：

    2026年1月8日

查看全文
  • Anish Sarkar,
  • Utpal Nandi,
  • Santanu Koley,
  • Chiranjit Changdar,
  • Bachchu Paul,
  • Partha Chowdhuri,
  • Pabitra Pal

摘要

Hyperspectral images (HSIs), which comprise numerous redundant spectral bands, are the most prevalent remote sensing sources for interpreting objects based on spectral band data. In the classification process, choosing a subset of spectral bands for data dimensionality reduction is known as band selection (BS). Deep learning (DL) based model with an attention mechanism can be used for BS, considering the nonlinear and global interaction among spectral bands. However, the existing DL based BS approaches using attention mechanisms are either unable to record both the spectral and spatial long-range information or depend only on queries, but not on keys and values. Moreover, the used reconstruction network (RecNet) in most of the studied BS techniques cannot recognize the features of images in compound scales since the network uses single-size kernels in convolution operations. To reduce these limitations of existing DL based BS approaches with attention mechanism, a novel DL based Hyperspectral BS (PSAA-MSRecNet) model consisting of position-sensitive axial attention (PSAA) with a multi-scale RecNet (MSRecNet) has been proposed. It takes advantage of a PSAA module that adds not only queries but also keys and values-dependent positional bias terms. The MSRecNet is employed after the PSAA module as a RecNet to search the image features in divergent scales. The proposed approach is able to effectively capture quality feature representations and, as a consequence, pick the most informative bands with higher accuracy in classification tasks than prior studied BS approaches, the majority of cases accross three standard HSI datasets.