Scholarlink

<p>Mural painting pigments are frequently applied in mixed form. Their reflectance spectra exhibit significant nonlinear characteristics due to multiple scattering effects, which poses substantial challenges for hyperspectral unmixing. This study proposes an extended multilinear mixing model-based convolutional autoencoder unmixing method (EMLM-CAE), which explicitly embeds the EMLM physical model into the decoder to characterize band-dependent nonlinear spectral variation while jointly capturing spatial and spectral information. Laboratory mixed-pigment experiments confirm significant band-dependent nonlinear features and show that the proposed method outperforms representative linear and nonlinear models in spectral reconstruction and abundance estimation. Validation on in situ hyperspectral images of Potala Palace mural paintings further demonstrates stable abundance inversion, achieving a reconstruction error of 0.0279 and a spectral angle distance of 0.0318. The results identify major pigments, including azurite, emerald green, cinnabar, minium, orpiment, and gold, and reveal widespread blue-green mixing with regionally distinct proportions.</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 nonlinear spectral unmixing method for mural painting pigments based on the extended multilinear mixing model and convolutional autoencoder

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月27日

查看全文
  • Wenliang Chen,
  • Xia Zhang,
  • Yi Hou,
  • Xiaoxuan Pan,
  • Feng Gao,
  • Haoran Ran,
  • Weihao Wang,
  • Youxin Sun

摘要

Mural painting pigments are frequently applied in mixed form. Their reflectance spectra exhibit significant nonlinear characteristics due to multiple scattering effects, which poses substantial challenges for hyperspectral unmixing. This study proposes an extended multilinear mixing model-based convolutional autoencoder unmixing method (EMLM-CAE), which explicitly embeds the EMLM physical model into the decoder to characterize band-dependent nonlinear spectral variation while jointly capturing spatial and spectral information. Laboratory mixed-pigment experiments confirm significant band-dependent nonlinear features and show that the proposed method outperforms representative linear and nonlinear models in spectral reconstruction and abundance estimation. Validation on in situ hyperspectral images of Potala Palace mural paintings further demonstrates stable abundance inversion, achieving a reconstruction error of 0.0279 and a spectral angle distance of 0.0318. The results identify major pigments, including azurite, emerald green, cinnabar, minium, orpiment, and gold, and reveal widespread blue-green mixing with regionally distinct proportions.