Physics-Guided Regime Unmixing
Pixel-wise nonlinear mixing beats fixed-regime models by 10%+ accuracy
Traditional Linear Mixing Model (LMM) dominates spectral unmixing but fails under multiple scattering. Existing nonlinear models apply a fixed nonlinear regime across entire scenes, ignoring spatial variability in scattering physics. In a new arXiv paper (2605.04247), Paula Pacheco and colleagues introduce Physics-Guided Regime Unmixing (PGRU), which estimates a pixel-wise scalar ξ_i ∈ [0,1] from observable physical features. This scalar activates nonlinear mixing only where justified, avoiding unnecessary complexity in linear regions. PGRU then combines residuals from three nonlinear models—Generalized Bilinear Model (GBM), Post-Nonlinear Mixing Model (PPNM), and Hapke—via learned attention, producing interpretable regime maps that show which model contributes most per pixel.
Experiments on three benchmark hyperspectral datasets (Samson, Jasper Ridge, Urban) show PGRU consistently improves unmixing accuracy over all baselines, with physical coherence (ρ) exceeding 0.90. The pixel-wise approach allows the model to adapt to heterogeneous scenes, such as urban areas with both linear (asphalt) and nonlinear (vegetation) mixing regimes. This work has direct implications for remote sensing applications in agriculture, mineral exploration, and environmental monitoring, where accurate material abundance estimation is critical.
- PGRU estimates pixel-wise scalar ξ_i ∈ [0,1] to dynamically activate nonlinear mixing only where needed
- Combines GBM, PPNM, and Hapke residuals via learned attention for interpretable regime maps
- Achieves physical coherence ρ > 0.90 on Samson, Jasper Ridge, and Urban datasets, outperforming fixed-regime models
Why It Matters
Enables more accurate land cover classification and material abundance estimation in complex, heterogeneous scenes.