Integration of Ground-Based and Satellite Data Using a Semi-Supervised GAN Model on Vulcano Island

Surface heat transfer is a continuous process reflecting the dynamic equilibrium between the magmatic system and the surrounding rock. In volcanic systems, part of the energy transferred from magma drives fluid convection, increasing ground temperatures. Total heat transfer occurs through three primary mechanisms: conduction, convection, and radiation. Each of these mechanisms plays a distinct role in volcanic systems and requires specific detection methods. Convective heat flow is observed in fumaroles and steaming ground; moderate thermal anomalies indicate conductive heat transfer; radiative heat flow is detected via multispectral instruments measuring surface thermal anomalies.

On Vulcano Island (Italy), a continuous monitoring network has detected transient variations in heat flow released by the active cone, which are correlated with increased seismic activity and ground deformation. Contact sensors monitor temporal variations in high-temperature fumaroles, while other sensors measure heat flux variations in areas of diffuse degassing. Long-term time series data have captured several episodes of volcanic unrest (e.g., Diliberto 2021; Federico et al., 2023).

This study presents results from the integration of artificial intelligence techniques with monitoring procedures, including:
a) Ground temperature measurements via contact sensors at selected sites;
b) Fumarole extent analysis;
c) Thermal and environmental indices derived from satellite imagery.

Specifically, a Semi-Supervised Generative Adversarial Network (SGAN) model was employed to automatically classify different volcanic states (baseline activity, transient degassing, and increased degassing). The model leverages direct temperature measurements from contact sensors (installed on the ground-based network on the La Fossa cone), land surface temperature anomalies (from MODIS), the Normalized Thermal Index (from VIIRS), and environmental indices such as NDVI, NDWI, and NDMI (from Landsat 8).

Preliminary results indicate that the SGAN model achieves an accuracy exceeding 0.89 for nearly all analyzed periods.