SpectralUnmix

SpectralUnmix is an R package for smooth non-negative spectral unmixing of hyperspectral datasets and astronomical IFU cubes.

Model

The package solves

\[ X \approx A S\]

where X is a spaxel-by-wavelength matrix, A contains spatial abundances, and S contains component spectra.

Each input spectrum is modeled as a non-negative combination of a small number of latent spectral components. ## Output example

The figure below shows the Coelho-based benchmark used in the validation page. The comparison is made on a controlled mixture benchmark built from the bundled stellar spectra.

Reproducible example

library(SpectralUnmix)

demo <- coelho_demo_spectra()

fit <- spectral_unmix(
  demo$matrix,
  k = 4,
  lambda_smooth = 0,
  niter = 400,
  lr = 0.03
)

plot(fit, type = "spectra", wavelength = demo$wavelength)
plot_reconstruction(fit, demo$matrix, wavelength = demo$wavelength)

Main functions

• spectral_unmix() fits the model.
• basis() and coef() provide NMF-style accessors.
• fitted(), predict(), and residuals() provide standard model methods.
• plot() and plot_reconstruction() provide quick diagnostics.
• cube_to_matrix() and matrix_to_cube() remain available for IFU workflows.

Installation

# install.packages("remotes")
remotes::install_github("RafaelSdeSouza/SpectralUnmix")

torch must be available in the local R installation.