AI-driven volcanic SO2 estimates using MSG-SEVIRI and Sentinel-5P TROPOMI

Abstract Sulfur dioxide (SO 2 ) is, after H 2 O and CO 2 , the most abundant volcanic gas released during shallow magma degassing and is readily detectable and quantifiable by remote sensing. Monitoring SO 2 provides key information on magma ascent rates, conduit dynamics, and eruption style and intensity, thereby supporting volcano monitoring and hazard assessment. The TROPOMI instrument onboard the low Earth Orbit Sentinel-5 Precursor delivers highly accurate and sensitive SO 2 retrievals but with limited temporal coverage (typically one overpass per day). In contrast, SEVIRI, onboard the geostationary MSG satellite, offers continuous high-frequency observations every 5–15 min, enabling real-time tracking of plume evolution. To further improve SO 2 detection capabilities, this study combines TROPOMI products with data from the MSG‑SEVIRI radiometer, by transferring TROPOMI’s retrieval accuracy to SEVIRI’s high-temporal-resolution observations. To this purpose, a data-driven AI model was implemented to estimate SO 2 vertical column densities (VCDs) at SEVIRI spatial and temporal resolution, using TROPOMI observations as reference. In this study, SO 2 VCDs were retrieved from Sentinel-5P TROPOMI Level 2 Offline DOAS measurements using slant column densities, averaging kernels, and plume height-adapted air mass factors. A multilayer perceptron was designed with an input layer receiving all available spectral bands from SEVIRI and a single linear output neuron, corresponding to the SO 2 VCD measured by TROPOMI. This approach enables SEVIRI data to inherit the sensitivity of TROPOMI while preserving SEVIRI’s native high-temporal-resolution and dense spatial sampling. Mount Etna (Italy), a persistently degassing open-conduit volcano, was selected as a test case. The trained network also allowed SO 2 retrieval from SEVIRI imagery acquired before Sentinel-5P’s launch, enabling the reconstruction of long-term degassing trends. SO 2 fluxes were computed and quantitatively validated against ground-based monitoring data. This integrated technique provides an effective tool for rapid and reliable volcanic hazard assessment, improving current retrieval methods and enhancing early-warning capability for aviation safety and climate studies.