EUVIMEDEUVIMED entdecken
Crossref

Hyperspectral unmixing with spectral variability: Reducing sensitivity to learning parameters, by combining linear and nonlinear NMF algorithms

Yannick Deville, Guillaume Faury

Digital Signal Processing · 2025-08 · 163 · 105241

DOI: 10.1016/j.dsp.2025.105241 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Digital Signal Processing, 2025-08. Verfasst von Yannick Deville und 1 weiteren Autor:innen. Die Arbeit wurde bisher 8-mal zitiert (Crossref) und nennt 43 Literaturangaben. Gefördert von: French National Research Agency.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Unsupervised hyperspectral unmixing methods aim at extracting the spectra of pure materials that are present in a hyperspectral image. A current major challenge consists of extending these methods to the case when each class of pure materials involves spectra that have a significant variability over image pixels (called spectral variability or intraclass variability). Un-mixing may essentially be reformulated as a matrix factorization problem, especially with nonnegativity constraints. Therefore, an attractive class of methods to solve this problem is (constrained) Nonnegative Matrix Factorization (NMF), which is a type of unsupervised machine learning methods. Whereas NMF results provided in the literature are most often limited to a single value of a performance parameter (SAM of estimated spectra), in the present paper we first analyze the sensitivity of performance to the selected adaptation (i.e. learning) gain and number of learning iterations. Due to the quite significant sensitivity thus observed for basic algorithms, it is difficult to automatically select their values of the above parameters, and overfitting may occur during learning. We solve these problems by introducing “hybrid NMF methods”. They essentially consist of successively running several types of NMF methods (respectively intended for linear and nonlinear mixtures), with different but compatible spaces for their adaptive variables, with different cost functions, and with part of the adaptive variables of each method initialized with the final values provided by the previously executed method for its corresponding variables. These hybrid methods have a low sensitivity to the adaptation gain and to the number of learning iterations, and they avoid overfitting over a large range of values of these parameters.

Förderung

  • French National Research Agency

Zitieren

APA

Deville, Y., & Faury, G. (2025). Hyperspectral unmixing with spectral variability: Reducing sensitivity to learning parameters, by combining linear and nonlinear NMF algorithms. Digital Signal Processing, 163, 105241. https://doi.org/10.1016/j.dsp.2025.105241

Vancouver

Deville Y, Faury G. Hyperspectral unmixing with spectral variability: Reducing sensitivity to learning parameters, by combining linear and nonlinear NMF algorithms. Digital Signal Processing. 2025;163:105241. doi:10.1016/j.dsp.2025.105241

BibTeX
@article{Deville2025,
  author = {Deville, Yannick and Faury, Guillaume},
  title = {Hyperspectral unmixing with spectral variability: Reducing sensitivity to learning parameters, by combining linear and nonlinear NMF algorithms},
  journal = {Digital Signal Processing},
  year = {2025},
  volume = {163},
  pages = {105241},
  doi = {10.1016/j.dsp.2025.105241}
}
Kostenlos starten →