Technical Summary: Physics-Guided Regime Unmixing (PGRU)

Problem Statement

Hyperspectral imaging captures surface reflectance across hundreds of bands, enabling material characterization for applications in agriculture, geology, and environmental monitoring. However, due to coarse spatial resolution relative to surface heterogeneity, individual pixels often contain mixed spectral responses from multiple materials. Spectral unmixing aims to recover pure material signatures (endmembers) and their fractional abundances.

The Linear Mixing Model (LMM) is the standard approach due to its simplicity and physical interpretability. However, LMM assumes photons interact with a single material before reaching the sensor, an assumption that fails under conditions of multiple scattering. While nonlinear extensions such as the Generalized Bilinear Model (GBM), the Post-Nonlinear Mixing Model (PPNM), and the Hapke radiative transfer model address this, they typically apply nonlinearity uniformly across entire scenes. This uniform application fails to distinguish which specific pixels require nonlinear treatment, often leading to suboptimal performance. Conversely, deep learning approaches (e.g., autoencoders) achieve strong reconstruction but operate as "black boxes," offering no physical explanation for regime selection or the conditions under which nonlinearity is active.

Methodology

The authors propose Physics-Guided Regime Unmixing (PGRU), a framework that estimates a pixel-wise scalar $\xi_i \in [0, 1]$ to activate nonlinear mixing only where physically justified. Unlike previous methods, PGRU does not require labeled data indicating whether pixels are linear or nonlinear; instead, regime assignment emerges from the interplay between reconstruction improvement and physical feature consistency.

Model Formulation

Each pixel $i$ is reconstructed as:
$$\hat{y}_i = S_{lin,i} + \xi_i \delta_{nl,i}$$
Where:

• $S_{lin,i}$ is the standard LMM component (non-negative abundances summing to one).
• $\delta_{nl,i}$ is a nonlinear residual.
• $\xi_i$ acts as a switch: $\xi_i \approx 0$ reduces the model to LMM, while $\xi_i \approx 1$ fully activates the nonlinear contribution.

Nonlinear Residual Construction

The residual $\delta_{nl,i}$ is a weighted combination of residuals from three physical models:
$$\delta_{nl,i} = \sum_{k} \alpha_{ik} \delta_{ik}$$
The weights $\alpha_{ik}$ are learned attention coefficients obtained via a temperature-controlled softmax, combining:

1. GBM: Captures pairwise scattering between endmembers.
2. PPNM: Introduces quadratic distortion over the linear mixture.
3. Hapke: Describes multiple scattering based on radiative transfer theory.
   Entropy regularization on $\alpha_{ik}$ encourages concentration toward the dominant physical mechanism at each pixel.

Physics-Guided Regime Estimation

The core innovation is that $\xi_i$ is guided by observable scene properties rather than solely by reconstruction error. It is computed as:
$$\xi_i = \sigma \left( \sum_{k} w_k F_{ik} + b \right)$$
Where:

• $F_{ik}$ represents physical and geometric features of pixel $i$ (spectral curvature, NDVI, NDVI gradient, EMP, DMP, LBP).
• $w_k$ are learned weights.
• $\sigma$ is the sigmoid function.
  This formulation is equivalent to logistic regression over the feature space, making the regime decision interpretable as a readable sum of feature contributions.

Objective Function

The model is trained using a per-pixel reconstruction gain $\Delta_{res,i} = \|y_i - S_{lin,i}\| - \|y_i - \hat{y}_i\|$. The full objective function $L$ is:
$$L = \sum_{i} \left[ -\tanh(\Delta_{res,i}) + \lambda_{feat}(\xi_i - \xi_{prior,i})^2 + \lambda_{sp} \xi_i \nabla^2 \xi_i \right] + \lambda_w \|w\|^2$$

• The first term rewards nonlinear activation where it reduces reconstruction error (saturated by $\tanh$).
• The second term anchors $\xi_i$ to the feature-based prior, with $\lambda_{feat}$ annealed during training.
• The third term promotes spatial coherence of the regime map via a Laplacian regularizer.
• The final term is L2 regularization on weights.
  No pixel labels are used; the regime assignment is entirely unsupervised.

Key Results

The authors evaluated PGRU on three benchmark datasets: Samson, Jasper Ridge, and Urban. Baselines included LMM, GBM, and PPNM.

• Reconstruction Accuracy: PGRU achieved consistent improvements over all baselines across all datasets.
  • Samson: rRMSE dropped from 0.495 (LMM) to 0.052 (PGRU).
  • Jasper Ridge: rRMSE dropped from 0.347 (LMM) to 0.068 (PGRU).
  • Urban: rRMSE dropped from 0.331 (LMM) to 0.064 (PGRU).
  • Notably, uniform nonlinear baselines (GBM, PPNM) did not consistently outperform LMM, confirming that uniform nonlinearity can be counterproductive.
• Physical Coherence: The correlation $\rho$ between the activation map $\xi$ and the reconstruction gain $\Delta_{res}$ exceeded 0.90 across all datasets (0.93, 0.91, and 0.92 respectively), confirming that nonlinear mixing is activated precisely where reconstruction improvements are observed.
• Interpretability: Feature weight analysis revealed scene-adaptive behavior:
  • In Jasper Ridge, NDVI dominated, concentrating nonlinear activation in dense vegetation (expected multiple scattering) and suppressing it over water.
  • In Urban, spectral curvature drove selection, reflecting structural complexity at material boundaries rather than vegetation interactions.
  • Spatial maps (e.g., Fig. 1) demonstrated that the model anchors regime selection to measurable scene properties, providing pixel-level explanations for why a pixel is nonlinear.

Significance and Claims

The paper claims that PGRU addresses a critical gap in spectral unmixing: determining why a specific pixel requires a nonlinear model. By grounding the regime scalar $\xi_i$ in observable physical features, the model produces interpretable regime maps without labeled data.

The authors assert that selective, explainable nonlinearity outperforms both uniform linear and uniform nonlinear approaches. They position PGRU as a practical alternative to globally linear or globally nonlinear models, offering a middle ground where nonlinearity is applied only where justified by physical evidence.

The authors maintain a modest tone regarding their claims, noting that:

1. These results are preliminary and require broader validation against recent deep-learning baselines.
2. The relative importance of the physical model ensemble, the feature-guided activation, and spatial regularization has not yet been disentangled through a systematic ablation study.
3. Future work is required to extend the method to blind unmixing settings, scenarios with ground-truth nonlinear interaction labels, and more complex mixing scenarios.