Abstract
Hyperspectral imagery collected from airborne or satellite sources inevitably suffers from spectral variability, making it difficult for spectral unmixing to accurately estimate abundance maps. The classical linear mixing model (LMM) generally fails to handle this issue effectively. We propose a novel spectral mixture model, the augmented linear mixing model (ALMM), to address spectral variability via a data-driven learning strategy. The ALMM separately models scaling factors (driven by illumination and topography) and other complex spectral variabilities (driven by environmental conditions and sensor noise) by introducing a spectral variability dictionary. By embedding a low-coherence dictionary learning technique into the unmixing framework, the model simultaneously learns the variability dictionary and estimates abundance maps. Extensive experiments demonstrate the superiority of ALMM over state-of-the-art methods.
Introduction
Hyperspectral sensors capture data across hundreds of narrow spectral bands, but their relatively low spatial resolution often results in "mixed pixels"—single pixels containing multiple materials, known as endmembers. Hyperspectral unmixing is the process of decomposing these pixels into their constituent endmembers and their corresponding proportions, or abundances.
The standard Linear Mixing Model (LMM) assumes that a material's spectral signature is constant across an entire image. However, in real-world scenarios, signatures vary due to illumination changes, topography, atmospheric conditions, and sensor noise—a phenomenon known as spectral variability. While previous models like the Extended LMM (ELMM) handle scaling and the Perturbed LMM (PLMM) handle additive noise, they often fail to capture the full complexity of real-world variability, leading to inaccurate abundance estimations.
The Augmented Linear Mixing Model (ALMM)
The proposed ALMM decomposes the observed signal into three primary components to better account for the diverse sources of spectral variability:
- Scaling Component ($AXS$): This component uses the known endmember dictionary ($A$) and applies pixel-wise scaling factors ($S$) to account for variations in topography and lighting.
- Augmented Variability Component ($EB$): This is the core novelty. It introduces a "spectral variability dictionary" ($E$) and a coefficient matrix ($B$) to capture complex residuals, such as atmospheric effects and nonlinear mixing, that simple scaling cannot explain.
- Residual Noise ($R$): Standard additive noise.
To solve this, the authors employ a data-driven dictionary learning strategy. A "low-coherence" prior is introduced to ensure that the atoms in the variability dictionary ($E$) remain distinct from the material signatures in $A$. This prevents the model from misidentifying a real material as a variation of another. The system is optimized using the Alternating Direction Method of Multipliers (ADMM).
Key Insights: The Nature of Variability
A critical observation of this work is that spectral variability in real scenes is not simply Gaussian noise. Analysis of real datasets shows that variability follows a complex, multi-modal distribution, often resembling a Gaussian Mixture Model rather than a simple bell curve.
By modeling scaling (physics-based) and other variabilities (data-driven) separately, the ALMM maintains physical interpretability while remaining flexible enough to handle environmental noise. The low-coherence constraint is essential, ensuring the variability term does not absorb signal belonging to distinct materials.
Experimental Results
The ALMM was evaluated against state-of-the-art methods like SUnSAL, PLMM, and ELMM across three datasets:
- Synthetic Data: ALMM achieved an abundance Root Mean Square Error (aRMSE) of 2.15, significantly lower than ELMM (3.23) and PLMM (6.21).
- Urban Dataset (HYDICE): In classification-based evaluations, ALMM reached an Overall Accuracy (OA) of 78.70%, a substantial improvement over ELMM’s 62.41%.
- Cuprite Dataset (AVIRIS): ALMM demonstrated a superior ability to distinguish between highly similar mineral signatures, producing "purer" abundance maps with higher contrast.
Conclusion
The Augmented Linear Mixing Model provides a robust framework for hyperspectral unmixing by simultaneously modeling illumination-driven scaling and complex environmental variations. While the optimization problem is non-convex and computationally more intensive than simpler models, the gains in accuracy and the ability to handle real-world spectral variability make it a significant advancement in the field. Future work will explore advanced spatial priors and automated selection of dictionary sizes.
Summary updated