Augmented Linear Mixing Model Overview

• ALMM is a spectral mixture model that explicitly models both scaling effects and non-scaling spectral variabilities to enhance material abundance estimation.
• It incorporates a low-coherence spectral variability dictionary alongside alternating ADMM optimization to efficiently decouple illumination and nonlinear effects.
• Empirical evaluations on synthetic and real datasets demonstrate significant improvements in accuracy, validating ALMM's effectiveness under challenging conditions.

The Augmented Linear Mixing Model (ALMM) is a spectral mixture model for hyperspectral unmixing, specifically designed to address the challenge of spectral variability in remotely sensed data. Unlike the classical Linear Mixing Model (LMM) and its prior extensions, the ALMM explicitly models both primary spectral variability caused by illumination/topography and secondary variability arising from environmental, instrumental, and nonlinear effects. ALMM couples parametric and data-driven strategies, introducing a low-coherence spectral variability dictionary and corresponding optimization framework for robust abundance estimation and dictionary learning (Hong et al., 2018).

1. Motivation and Background

Hyperspectral unmixing is the estimation of material abundances from mixed spectral pixels under the presence of spectral variability. The classic Linear Mixing Model (LMM) assumes that each observed spectrum $y_n \in \mathbb{R}^D$ is a convex combination of $P$ endmember spectra:

$y_n = M a_n + e_n$

subject to abundance nonnegativity (ANC: $a_n \geq 0$) and sum-to-one constraint (ASC: $1^\top a_n = 1$). However, LMM does not account for inherent variability in endmember spectra due to illumination, local atmospheric conditions, instrument artifacts, or nonlinear mixing, causing significant unmixing errors in practice.

Prior attempts to extend LMM include the Extended LMM (ELMM), which models endmember-wise scaling but neglects other variability modes, and the Perturbed LMM (PLMM), which uses additive perturbations without separating scaling effects from complex variabilities. The ALMM is developed to provide a unified model for both scaling and residual spectral variability (Hong et al., 2018).

2. Model Formulation

The ALMM expresses each observed pixel spectrum as:

$y_n = s_n M a_n + D b_n + e_n$

where:

• $s_n \geq 0$ is a scalar scaling factor per pixel (illumination/topography),
• $M \in \mathbb{R}^{D \times P}$ is the fixed endmember dictionary,
• $a_n \in \mathbb{R}^P$ is the abundance vector ($a_n \geq 0$, $P$0),
• $P$1 is a data-driven spectral variability dictionary,
• $P$2 are variability coefficients per pixel,
• $P$3 is additive noise.

The term $P$4 models dominant per-pixel scaling (illumination/topography effects). The additive term $P$5 captures non-scaling spectral variabilities: environmental factors (e.g., atmospheric absorption), sensor artifacts, and nonlinear mixture residuals.

3. Regularization and Priors

Properly constraining ALMM’s expanded solution space is critical. The following priors and regularizations are imposed:

• Abundances $P$6:
  • ANC ($P$7) and ASC ($P$8).
  • Sparsity regularization: $P$9, with $y_n = M a_n + e_n$0.
• Variability coefficients $y_n = M a_n + e_n$1:
  • Unconstrained in sign; penalized via $y_n = M a_n + e_n$2, $y_n = M a_n + e_n$3.
• Dictionary $y_n = M a_n + e_n$4:
  • Low-coherence with endmember matrix $y_n = M a_n + e_n$5: penalize $y_n = M a_n + e_n$6 (encouraging near-orthogonality).
  • Self-incoherence/orthogonality: penalize $y_n = M a_n + e_n$7, $y_n = M a_n + e_n$8.
• Scaling factors $y_n = M a_n + e_n$9:
  • Nonnegativity ($a_n \geq 0$0).

The complete objective across $a_n \geq 0$1 pixels is:

$a_n \geq 0$2

subject to $a_n \geq 0$3, $a_n \geq 0$4, $a_n \geq 0$5.

4. Optimization Framework

Solving the ALMM objective is a nonconvex problem due to the coupling of variables. Optimization proceeds by block coordinate descent, alternating between:

• Spectral Unmixing (SU): Fix $a_n \geq 0$6; for each pixel update $a_n \geq 0$7. The SCLSU (Scaled Constrained Least Squares Unmixing) method is used to decouple $a_n \geq 0$8 and $a_n \geq 0$9, enabling nonnegativity and sum-to-one projection for abundance, and nonnegativity for scale.
• Spectral Variability Dictionary Learning (SVDL): Fix $1^\top a_n = 1$0; update $1^\top a_n = 1$1. Auxiliary variables enforce low-coherence and self-incoherence/orthogonality, and updates are performed via ADMM.

Both subproblems are convex. Alternating Direction Method of Multipliers (ADMM) enables efficient updates, with primal and dual residuals monitored for convergence. In practice, SU converges in 50-100 iterations and SVDL in 100-200, with per-iteration costs $1^\top a_n = 1$2 for SU and $1^\top a_n = 1$3 for SVDL.

5. Empirical Evaluation

ALMM’s performance was assessed on synthetic and real datasets (Urban—HYDICE and Cuprite—AVIRIS). Major experimental results include:

• Synthetic Data ($1^\top a_n = 1$4, $1^\top a_n = 1$5, $1^\top a_n = 1$6 endmembers, $1^\top a_n = 1$7 noise): For aRMSE, rRMSE, and aSAM, ALMM achieves $1^\top a_n = 1$8, $1^\top a_n = 1$9, and $y_n = s_n M a_n + D b_n + e_n$0 respectively, substantially outperforming FCLSU, CLSU, SCLSU, SUnSAL, SSUnSAL, ELMM, and PLMM.
• Urban Data ($y_n = s_n M a_n + D b_n + e_n$1, $y_n = s_n M a_n + D b_n + e_n$2, $y_n = s_n M a_n + D b_n + e_n$3): ALMM attains $y_n = s_n M a_n + D b_n + e_n$4 rRMSE, $y_n = s_n M a_n + D b_n + e_n$5 aSAM, and $y_n = s_n M a_n + D b_n + e_n$6 OA versus next-best ELMM at $y_n = s_n M a_n + D b_n + e_n$7, $y_n = s_n M a_n + D b_n + e_n$8, and $y_n = s_n M a_n + D b_n + e_n$9 respectively.
• Cuprite Data ($s_n \geq 0$0, $s_n \geq 0$1, $s_n \geq 0$2): On four principal minerals, ALMM achieves $s_n \geq 0$3 rRMSE, $s_n \geq 0$4 aSAM, and $s_n \geq 0$5 OA.

ALMM demonstrated robust performance for regularization parameters $s_n \geq 0$6, $s_n \geq 0$7, and dictionary size $s_n \geq 0$8. Notably, even at a low SNR of $s_n \geq 0$9, the abundance RMSE for ALMM remains below $M \in \mathbb{R}^{D \times P}$0, while all comparators exceed $M \in \mathbb{R}^{D \times P}$1 (Hong et al., 2018).

6. Implementation and Extensions

For practical implementation:

• Initialize $M \in \mathbb{R}^{D \times P}$2 via SCLSU ($M \in \mathbb{R}^{D \times P}$3, $M \in \mathbb{R}^{D \times P}$4); $M \in \mathbb{R}^{D \times P}$5 as a random orthonormal $M \in \mathbb{R}^{D \times P}$6 matrix.
• Select regularizations: $M \in \mathbb{R}^{D \times P}$7 in $M \in \mathbb{R}^{D \times P}$8; $M \in \mathbb{R}^{D \times P}$9 up to $a_n \in \mathbb{R}^P$0 larger.
• Choose $a_n \in \mathbb{R}^P$1 based on spectrum complexity, typically $a_n \in \mathbb{R}^P$2.
• Terminate ADMM when residuals fall below $a_n \in \mathbb{R}^P$3 or after $a_n \in \mathbb{R}^P$4 iterations.

Possible model extensions include spatial regularization (e.g., total variation or graph Laplacian on abundances), kernelized variants or deep autoencoder integration for nonlinear mixing, joint endmember–variability learning with coupled $a_n \in \mathbb{R}^P$5, and multi-scale ALMMs for region-specific variability (Hong et al., 2018).

7. Significance and Applications

The ALMM framework provides a principled approach to modeling and mitigating spectral variability, leading to improved hyperspectral unmixing accuracy under realistic conditions where illumination, topography, and extrinsic atmospheric/instrumental artifacts introduce complex spectral distortions. By jointly learning a low-coherence, orthogonal variability dictionary and abundance maps via alternating ADMM updates, ALMM outperforms classical and contemporary state-of-the-art methods, demonstrating resilience to strong spectral variability and noise. Applications include remote sensing, mineral mapping, agriculture, and environmental monitoring, wherever precise material quantification from hyperspectral imagery is essential (Hong et al., 2018).