A TIR-based surrogate model emulating radiative transfer for volcanic SO2 quantification

Volcanic sulfur dioxide (SO₂) is a primary indicator of magmatic degassing and eruptive activity and is routinely monitored using satellite observations in the Thermal Infrared (TIR) spectral range, including data from sensors such as the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER). Quantification of volcanic SO₂ from TIR measurements typically relies on radiative transfer models, such as MODTRAN, which are computationally expensive and limit their applicability in near-real-time monitoring scenarios.

In this study, we present a Neural Network-based surrogate modeling approach designed to emulate a physically based TIR radiative transfer model for volcanic SO₂ quantification from ASTER-like observations. Input features include TIR brightness temperatures, viewing geometry, and plume altitude.

The surrogate model is trained on a large synthetic dataset generated using MODTRAN simulations spanning a wide range of atmospheric, surface, and plume conditions, considering various eruptive scenarios. Validation results show that the surrogate accurately reproduces the MODTRAN-simulated radiances and the corresponding SO₂ column estimates, with errors well below the uncertainty associated with satellite noise and model assumptions.

By reducing computational costs by several orders of magnitude, the proposed surrogate enables efficient inversion of volcanic  SO₂ from ASTER TIR satellite data while preserving the physical consistency of the original radiative transfer model. This approach is particularly suited for operational volcanic monitoring, ensemble retrievals, and uncertainty propagation in  SO₂ quantification.