Hyperspectral Unmixing
Also known as: Spectral Mixture Analysis, Linear Spectral Unmixing, Blind Source Separation (Hyperspectral), Hiperspektral Ayrıştırma
Hyperspectral unmixing is a signal processing technique that decomposes each pixel of a hyperspectral image into a collection of pure material spectra (endmembers) and their corresponding fractional abundances. Because sensor resolution often causes multiple land-cover types to co-occupy a single pixel, unmixing recovers sub-pixel compositional information that conventional classification cannot. Keshava and Mustard (2002) provided the foundational signal-processing framework that unified prior geological and remote-sensing work under a rigorous linear mixture model.
Key highlights
- Recovers sub-pixel material proportions that hard classifiers cannot provide.
- Interpretable outputs: endmember spectra are physically meaningful and matchable against spectral libraries.
- Scales efficiently to large image cubes after initial dimensionality reduction.
- Provides a rigorous probabilistic interpretation when formulated under a Bayesian or constrained least-squares framework.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use hyperspectral unmixing when sensors capture scenes at a spatial resolution coarser than the scale of heterogeneity among materials of interest, such as minerals, vegetation types, or urban surfaces. The linear mixture model is appropriate when light interacts with materials in a single-bounce fashion (intimate mixtures require nonlinear extensions). Alternatives include pixel-based hard classifiers when spatial resolution is sufficient, or nonlinear unmixing methods when multiple-scattering dominates. The method requires that the number of endmembers be specified or estimated, and that at least quasi-pure pixel spectra exist in the scene.
Strengths & limitations
- Recovers sub-pixel material proportions that hard classifiers cannot provide.
- Interpretable outputs: endmember spectra are physically meaningful and matchable against spectral libraries.
- Scales efficiently to large image cubes after initial dimensionality reduction.
- Provides a rigorous probabilistic interpretation when formulated under a Bayesian or constrained least-squares framework.
- The linear mixture model fails in scenes with intimate mixing or significant multiple scattering.
- Performance degrades when no pure pixels exist for every endmember in the scene.
- The number of endmembers p must be set a priori or estimated, introducing sensitivity to misspecification.
- Spectral variability of endmembers across the scene (due to illumination, viewing angle, or phenology) violates the fixed-endmember assumption and inflates abundance errors.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
What is the difference between hard classification and spectral unmixing?
Hard classification assigns each pixel to exactly one class, assuming the pixel is spectrally pure. Spectral unmixing instead estimates a continuous mixture of materials within each pixel, yielding fractional abundance maps. Unmixing is appropriate when sensor pixels are larger than the spatial scale of the materials, whereas hard classification suffices when pixels are predominantly pure.
Is the linear mixture model always valid?
No. The linear model assumes single-bounce light interaction between the sensor and surface materials—a reasonable approximation for most flat, heterogeneous canopy or mineral surfaces. When photons scatter among multiple materials before reaching the sensor (intimate or intra-pixel mixing, dense vegetation), nonlinear mixture models such as bilinear or radiative-transfer-based approaches are more appropriate.
How do I determine the number of endmembers?
Common strategies include the Virtual Dimensionality (VD) estimator based on Neyman-Pearson detection theory, information-theoretic criteria such as minimum description length, and scree-plot inspection of PCA eigenvalues. Analysts should also consult prior knowledge of the study scene, as statistical estimators can be sensitive to noise level and scene complexity.
Sources
- 1.Keshava, N., & Mustard, J. F. (2002). Spectral unmixing. IEEE Signal Processing Magazine, 19(1), 44–57.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 2). Hyperspectral Unmixing. ScholarGate. https://scholargate.app/remote-sensing/hyperspectral-unmixing