Physics-informed neural networks with Sobolev shape regularization for thermal infrared temperature and emissivity separation: an open, leakage-controlled benchmark against the ASTER/HyTES TES algorithm
Year: 2026
Type: preprint
Abstract: Retrieving Land Surface Temperature (LST) and Land Surface Emissivity (LSE) from thermal infrared radiances is a classically ill-posed inverse problem (N channels for N + 1 physical unknowns). Operational algorithms, notably Temperature and Emissivity Separation (TES), resolve this Planck degeneracy using an empirical power law linking minimum emissivity to spectral contrast. While computationally expedient, this empirical heuristic amplifies channel noise into high-frequency sawtooth oscillations across hyperspectral channels, compounds atmospheric moisture errors, and degrades on anomalous mineralogies. We hypothesize that physics-informed neural networks regularized in function space by discrete Sobolev penalties and forward Planck consistency can resolve the Planck degeneracy without empirical contrast regressions. To test this hypothesis, we present a neural architecture featuring decoupled, thermodynamically bounded output heads (Ts > 200 K, ε in (0, 1)), discrete Sobolev slope (W1,1) and curvature (W2,1) shape regularizers, and forward radiative consistency feedback. Using the open OpenTIR-PINN benchmark suite (240 curated laboratory spectra across 11 material classes, 30 MODTRAN 5.2 profiles, and a 2 x 2 double-quarantine evaluation matrix), we conduct a paired evaluation against operational TES across hyperspectral (HyTES, 223 channels) and multispectral (ASTER, 5 bands) regimes. In the hyperspectral regime, relying solely on surface-leaving radiance without auxiliary atmospheric telemetry, the network decisively outperforms classical TES: it reduces LST root-mean-square error by 44.0% (1.186 vs. 2.119 K), cuts channel-averaged emissivity error by 68.4%, halves water-vapor error propagation, and eliminates the 18-fold high-frequency noise amplification of classical TES to replicate natural spectral smoothness. Contiguous spectral sampling provides sufficient structural redundancy to constrain the physical inverse manifold directly. On multispectral ASTER, by contrast, sparse five-band sampling causes mathematical underdetermination that cannot decouple surface emission from atmospheric attenuation without auxiliary telemetry; here, the network achieves broad operational parity with classical TES in temperature while reconstructing emissivity shapes with superior angular fidelity. Stress-testing on quarantined golden targets demonstrates that on anomalous mineralogies exhibiting extreme absorption depths (exemplified by kaolinite), the neural prior induces cold temperature biases (-4 to -12 K). Model ensemble spread captures these out-of-domain targets on HyTES, but fails on ASTER, establishing the necessity of radiance-derived novelty filters before satellite deployment. Finally, batched inference executes tens of thousands of pixels per second on standard CPU hardware without GPU acceleration, demonstrating that physics-informed neural inversion is computationally tractable and ready for next-generation satellite missions, including NASA's SBG-TIR and ESA's LSTM.
DOI: 10.31223/x5mj8j
Language: English
Subfield: Civil and Structural Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235550160
