Hyperspectral sensors capture the "fingerprint" of materials by measuring light across hundreds of narrow spectral bands. In an ideal world, every pixel in a satellite image would represent exactly one material—pure grass, pure water, or pure asphalt. In reality, because of limited spatial resolution, most pixels are "mixed," containing a weighted combination of several constituent materials, known as endmembers.
The goal of hyperspectral unmixing is to reverse this process: given a set of mixed pixels, identify the spectral signatures of the underlying endmembers and their relative proportions (abundances).
The Pure Pixel Assumption
For decades, the field relied on a convenient "cheat code" known as the Pure Pixel Assumption (PPA). Geometrically, if you plot hyperspectral data in a high-dimensional space, the linear mixing model suggests that all data points fall inside a simplex (a generalized triangle or tetrahedron) whose vertices are the endmembers.
As Xiao Fu 0001 and José M. Bioucas-Dias explain in Self-Dictionary Sparse Regression for Hyperspectral Unmixing, if at least one "pure" pixel exists for every material, the unmixing problem becomes a search for the extreme points of the data cloud. Algorithms like Vertex Component Analysis (VCA) simply hunt for these corners.
However, in complex urban environments or low-resolution imagery, this assumption frequently fails. If no pure pixels exist, the "corners" of the data cloud are rounded off, and traditional geometric methods end up identifying mixtures as if they were pure materials, leading to significant errors.
Joint Estimation and Bayesian Inference
When pure pixels are absent, we can no longer treat endmember extraction and abundance estimation as two separate steps. Instead, we must solve them simultaneously.
Nicolas Dobigeon and Alfred O. Hero pioneered a unified Bayesian framework for this in Joint Bayesian endmember extraction and linear unmixing for hyperspectral imagery. By treating both the endmembers and their proportions as unknown random variables, the model can "infer" where the true vertices of the simplex must be, even if no data points actually touch them. This statistical approach provides a buffer against noise and allows the model to hallucinate the pure signatures based on the distribution of the mixtures.
Archetypal Analysis as Robust NMF
A powerful alternative to pure geometric search is Archetypal Analysis, a specific form of Non-negative Matrix Factorization (NMF). In Non-negative Matrix Factorization via Archetypal Analysis, Andrea Montanari and Hamid Javadi describe a framework that relaxes the "separability" requirement.
Instead of forcing the endmembers to be actual observations from the dataset, Archetypal Analysis looks for "archetypes" that are convex combinations of the data points themselves. This creates a feedback loop: the archetypes must explain the data well, but the data must also justify the existence of the archetypes. This trade-off makes the decomposition stable even when the data is highly mixed.
Building on this, Alexandre Zouaoui and Jocelyn Chanussot introduced Entropic Descent Archetypal Analysis (EDAA) in Entropic Descent Archetypal Analysis for Blind Hyperspectral Unmixing. EDAA uses entropic gradient descent to navigate the complex optimization landscape of blind unmixing, outperforming deep learning methods by focusing on the inherent geometry of the "mixed-only" data.
Message Passing and Ellipsoidal Geometry
Other researchers have looked toward signal processing and advanced geometry to solve the "missing vertex" problem. In Hyperspectral Unmixing Via Turbo Bilinear Approximate Message Passing, Jeremy P. Vila and Philip Schniter apply message-passing algorithms to iteratively refine estimates of both endmembers and abundances. This "Turbo" approach allows the algorithm to exploit the statistical structure of the image to fill in the gaps where pure pixels are missing.
In specialized fields like Terahertz imaging, Chia-Hsiang Lin and colleagues proposed HYPERION in HYPERION: Hyperspectral Penetrating-type Ellipsoidal Reconstruction for Terahertz Blind Source Separation. Rather than searching for a simplex, they use ellipsoidal geometry to "penetrate" the mixed signals. This is particularly useful when the data is not just mixed, but also diluted or attenuated, as is common in non-destructive testing.
Unifying the Field
The shift away from the pure pixel assumption represents a transition from "searching" to "inferring." As summarized in the HySUPP Python Package overview by Behnood Rasti and others, the modern researcher now has a suite of tools—ranging from multiscale bundles to deep NMF—that treat the absence of pure pixels not as a failure condition, but as a standard feature of real-world remote sensing.