PLMM: Perturbed Linear Mixing Model

• The PLMM is a statistical framework that incorporates additive spectral perturbations to account for pixel-wise variability in hyperspectral imagery.
• It leverages ADMM optimization and convex penalty functions to enforce abundance smoothness and physical constraints for improved model interpretability.
• Empirical evaluations show that PLMM reduces reconstruction errors and improves abundance estimation compared to traditional linear mixing models.

The Perturbed Linear Mixing Model (PLMM) is a statistical framework for hyperspectral unmixing that addresses spatial and spectral variability in endmember signatures. In hyperspectral data analysis, each pixel is typically modeled as a linear mixture of reference spectral signatures ("endmembers") and their respective abundance fractions. Traditional linear mixing models neglect intra-class variability, which in practice can induce significant estimation error. The PLMM extends the standard model by introducing explicit spatially and spectrally dependent perturbations for each pixel-endmember pair, improving both fit and interpretability of abundance and endmember estimates in hyperspectral imagery (Thouvenin et al., 2015).

1. Model Formulation

Given a hyperspectral data matrix $\mathbf{X} \in \mathbb{R}^{L \times N}$ with $L$ spectral bands and $N$ pixels, the PLMM represents each pixel $\mathbf{x}_n$ as: $$\mathbf{x}_n = \sum_{k=1}^K a_{kn}\bigl(\mathbf{m}_k + \Delta\mathbf{m}_{n,k}\bigr) + \mathbf{n}_n, \quad n=1, \dots, N$$ where:

• $\mathbf{M}\in\mathbb{R}^{L\times K}$: matrix of $K$ reference endmembers (columns $\mathbf{m}_k$),
• $\mathbf{A}\in\mathbb{R}^{K\times N}$: abundance matrix (columns $\mathbf{a}_n = [a_{1n}, ..., a_{Kn}]^T$),
• $L$0: per-pixel spectral perturbation matrix (columns $L$1),
• $L$2: additive Gaussian noise.

The model enforces nonnegativity and sum-to-one constraints on abundances, as well as nonnegativity on the reference and perturbed endmembers: $L$3

PLMM introduces convex penalty functions to regularize the problem:

• Abundance smoothness: $L$4, where $L$5 encodes spatial differences.
• Endmember penalty: $L$6, e.g., distance to a reference library or mutual-distance/minimum-volume in projected space.
• Variability penalty: $L$7.

The joint optimization problem becomes: $L$8

2. Optimization via ADMM

Solving the PLMM unmixing problem requires minimizing a nontrivial, though block-convex, cost. The approach utilizes a block-coordinate descent (BCD) scheme, alternating updates between abundances $L$9, reference endmembers $N$0, and per-pixel perturbations $N$1. Each subproblem is convex and admits an efficient solution via the alternating direction method of multipliers (ADMM).

The augmented Lagrangian for a general constrained minimization is: $N$2 with iterative updates for primal and dual variables. Convergence is tracked by monitoring the primal and dual residuals until they fall below prespecified tolerances, with adaptive adjustment of the penalty parameter $N$3.

Specific subproblems in PLMM are updated as follows:

• Abundances $N$4: Each pixel's abundance vector is updated with nonnegativity and sum-to-one constraints, spatial smoothness, and splitting to facilitate ADMM.
• Reference endmembers $N$5: Updates are performed per spectral band, with constraints to ensure physical interpretability and flexibility in the penalty function.
• Perturbed endmember matrices $N$6: Each is updated per-pixel with a Frobenius penalty, ensuring nonnegativity of the resultant endmember.

Parameter settings for $N$7 and convergence thresholds are selected to ensure strongly convex subproblems and effective BCD convergence to a stationary point.

3. Regularization and Physical Constraints

The PLMM framework enforces several regularization and physical plausibility constraints to prevent degenerate or nonphysical solutions. Abundance smoothness exploits the spatial structure of hyperspectral images, penalizing sharp spatial changes unless dictated by data. Penalties on endmember matrices can anchor solutions close to known libraries or encourage diversity in the endmember set via mutual distances or minimum-volume constraints. The variability penalty $N$8 discourages excessive per-pixel deviations, providing control over the strength of modeled spectral variability.

These regularizations are critical due to the ill-posedness of unmixing with pixel-specific variability, and their proper tuning directly affects both reconstruction error and physical interpretability of the solution.

4. Empirical Evaluation and Performance

Empirical evaluation with synthetic and real datasets demonstrates the advantages of the PLMM compared to previous state-of-the-art algorithms such as VCA/FCLS, SISAL/FCLS, AEB, and FDNS. On synthetic datasets (128×64 pixels, 413 bands, SNR=30 dB, $N$9 or 6 endmembers), PLMM attains comparable or lower mean-squared errors (MSEs) for abundances and achieves substantially reduced reconstruction errors (down to $\mathbf{x}_n$0). It successfully recovers spatial maps of variability energy and remains robust even when pure pixels are absent.

On real hyperspectral images (AVIRIS Moffett Field and Cuprite), with $\mathbf{x}_n$1 ranging from approximately 3 to 10, PLMM consistently yields the lowest reconstruction errors observed (e.g., $\mathbf{x}_n$2 vs. $\mathbf{x}_n$3–$\mathbf{x}_n$4 for comparators). Variability maps generated by PLMM highlight zones suspected to be affected by nonlinear mixing phenomena—such as coastal pixels in Moffett and highly altered regions in Cuprite.

Computation time is greater than for classical LMM-based methods, ranging from tens to thousands of seconds for small images, but the implementation remains tractable for datasets up to several tens of thousands of pixels.

5. Comparison with Classical Linear Mixing Models

Traditional linear mixing models (LMMs) assume fixed spectral endmembers for all pixels, disregarding the pixel-to-pixel variability present in realistic scenarios. This can induce systematic errors in unmixing, especially when physical conditions vary spatially (e.g., illumination, atmospheric effects, material impurities). The PLMM explicitly models variability via additive perturbations, allowing for pixelwise adaptation of endmember signatures.

Empirical comparisons demonstrate that PLMM achieves improved data fitting, reduced abundance estimation errors, and the ability to explicitly estimate the spatial organization of spectral variability—capabilities not available in classical LMMs (Thouvenin et al., 2015).

6. Practical Implementation and Limitations

The PLMM + ADMM framework entails a computational burden due to the number of variables and required block-wise updates. Nevertheless, strong convexity in each block and efficient closed-form updates ensure practical convergence for hyperspectral images of moderate size. The method's flexibility accommodates various forms of prior information via penalty selection. A plausible implication is that for very large datasets or real-time scenarios, further optimization (e.g., parallelization, warm starts, or dimensionality reduction) may be beneficial.

A limitation is the increased complexity in model selection and parameter tuning; regularization strengths ($\mathbf{x}_n$5, $\mathbf{x}_n$6, $\mathbf{x}_n$7) and penalty forms influence estimation quality and computational efficiency.

7. Significance and Applications

The PLMM provides a flexible and physically interpretable approach for hyperspectral unmixing in the presence of spectral variability. It yields explicit per-pixel variability estimates, enhances fidelity of abundance and endmember estimation, and offers new capabilities for diagnosing nonlinear mixing regions in complex scenes. Its utility has been demonstrated on both synthetic benchmarks and real AVIRIS images, establishing its relevance in remote sensing and spectral analysis where endmember variability is the norm (Thouvenin et al., 2015).