Physics-Augmented Domain Adaptation for Hyperspectral Image Classification Under Material Degradation

Contemporary spaceborne hyperspectral imaging platforms such as PRISMA, DESIS, and EnMAP achieve 94–99% classification accuracy on pristine laboratory scenes. However, a critical operational barrier persists: classification performance degrades systematically when imaged materials experience environmental stress including atmospheric weathering, UV oxidation, thermal cycling, corrosion, and space radiation exposure. Classifiers trained exclusively on pristine spectra fail catastrophically on degraded targets, with documented accuracy drops of 20–40 percentage points. Current domain adaptation methods such as DANN, CORAL, and DIAT assume distribution shifts arise from sensor differences or illumination variation, but environmental material degradation is fundamentally physics-governed, following deterministic kinetic laws such as exponential or power-law decay rather than arbitrary statistical shifts. This work presents Physics-Augmented Domain Adaptation (PADA), which explicitly encodes exponential reflectance decay kinetics from NASA’s Long Duration Exposure Facility (LDEF) mission—5.76 years of space exposure with 10,000 material samples—into domain adaptation via a learned regularization term. The physics loss constrains learned features to respect exponential decay kinematics, validated against measured LDEF spectral data with R²=0.91 and RMSE=0.087. On terrestrial hyperspectral benchmarks under synthetic heavy degradation simulating 5.76-year equivalent exposure, PADA achieves 79.2% accuracy versus 69.8% baseline (DANN), yielding +9.4 percentage point improvement. Isolated physics contribution through ablation study yields +9.9 percentage points with statistical significance demonstrated through Wilcoxon signed-rank test (p=0.0053 Bonferroni-corrected, effect size r=0.52). Bootstrap 95% confidence interval spans [7.8%, 11.0%] based on 5-fold stratified cross-validation with 5 seeds per fold (N=25 configurations, power=78%). With Monte Carlo Dropout uncertainty quantification, PADA achieves 84.1% accuracy (+14.3 percentage point cumulative gain) and enables selective deployment with 89.3% accuracy on high-confidence predictions retaining 65% of data (ROC-AUC=0.87 calibration validation). This Phase 1 work validates proof-of-concept methodology using controlled synthetic LDEF-parameterized degradation on terrestrial hyperspectral proxies and does NOT claim operational deployment readiness on real space-exposed or weathered materials. Phase 2 development (12–18 months) will address cross-domain validation, multi-material physics parameterization, and real-world applicability with specific milestone deliverables.