Scholarlink

<p>Satellite images are very often affected by noise that degrades both spectral fidelity and geological structure. This research proposes two Neural Cellular Automata (NCA) based denoising methods - Convolutional Neural Cellular Automata (CNN NCA), and Graph Neural Cellular Automata (GNN NCA) that extends NCA from standard grayscale (or RGB) images to hyperspectral cubes. While CNN NCA computes residual NCA updates using convolutional feature extraction, GNN NCA uses graph attention-based neighbourhood aggregation to preserve local structure during iterative updates. To evaluate the performance of the proposed NCA based denoising techniques they were experimented on real hyperspectral patches from Chandrayaan 1 Moon Mineralogy Mapper (M<sup>3</sup>, 85 bands) and Chandrayaan 2 Imaging IR Spectrometer (IIRS, 256 bands), with Gaussian and salt and pepper noise injection for robustness analysis. GNN-NCA improves SSIM by <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(15.29\%\)</EquationSource> </InlineEquation> and reduces MSE by <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(3.17\%\)</EquationSource> </InlineEquation> on Ch-1 M<sup>3</sup> data, and improves SSIM by <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(21.69\%\)</EquationSource> </InlineEquation> and reduces MSE by <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(6.82\%\)</EquationSource> </InlineEquation> on Ch-2 IIRS data, relative to the corresponding noise-added inputs, indicating stronger structural preservation and error suppression.</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

Neural cellular automata for hyperspectral denoising of Chandrayaan-1 M3 and Chandrayaan-2 IIRS sensor data from the Manzius U–Boguslawsky M Region with convolutional and graph neural updates

  • 期刊论文

  • 发表日期：

    2026年7月25日

查看全文
  • M. Gayathri Lakshmi,
  • Anurag Dutta

摘要

Satellite images are very often affected by noise that degrades both spectral fidelity and geological structure. This research proposes two Neural Cellular Automata (NCA) based denoising methods - Convolutional Neural Cellular Automata (CNN NCA), and Graph Neural Cellular Automata (GNN NCA) that extends NCA from standard grayscale (or RGB) images to hyperspectral cubes. While CNN NCA computes residual NCA updates using convolutional feature extraction, GNN NCA uses graph attention-based neighbourhood aggregation to preserve local structure during iterative updates. To evaluate the performance of the proposed NCA based denoising techniques they were experimented on real hyperspectral patches from Chandrayaan 1 Moon Mineralogy Mapper (M3, 85 bands) and Chandrayaan 2 Imaging IR Spectrometer (IIRS, 256 bands), with Gaussian and salt and pepper noise injection for robustness analysis. GNN-NCA improves SSIM by \(15.29\%\) and reduces MSE by \(3.17\%\) on Ch-1 M3 data, and improves SSIM by \(21.69\%\) and reduces MSE by \(6.82\%\) on Ch-2 IIRS data, relative to the corresponding noise-added inputs, indicating stronger structural preservation and error suppression.