Surface heat transfer is a continuous process that describes the dynamical equilibrium between the magmatic system and the host rock. In volcanic systems, part of the energy transfer from magma drives fluid convection and increases ground temperatures. Total heat transfer occurs by the combination of three processes: conductive, convective and radiative heat flow. Any of these processes has a different load in the volcanic system, and their detection needs different methodologies and approaches. Steaming ground and fumaroles show the convective heat flow reaching the ground surface; mild thermal anomalies reflect conductive heat transfer within the ground; and multi-spectral instruments can detect the heat flow radiating from the ground surface. The continuous monitoring network deployed on the Island of Vulcano (Italy) showed transient variations in the heat flow released by the active cone, always related to increasing seismic activity and ground deformations. Some contact sensors monitor the time variations of high-temperature fumaroles; other sensors monitor time variations of heat flux in the mild thermal zone associated with diffuse degassing. The resulting long-term time series tracked several unrests (e.g. Diliberto 2021, Federico et al., 2023).

We present results from the integration of AI techniques and different monitoring procedures, measuring the following: a) ground temperature by contact sensors on selected sites; b) fumarole extension; c) thermal and environmental indices from satellite imagery. In particular, we employed a Semi-Supervised Generative Adversarial Network (SGAN) model to classify different levels of volcanic states (background activity, transient degassing and updated degassing level) automatically. The model leverages direct temperature measurements from contact sensors (deployed by the ground-based network on La Fossa cone), land surface temperature anomalies (from MODIS), the Normalized Thermal Index (from VIIRS) and the environmental indices NDVI, NDWI, NDMI (from Landsat 8). Preliminary results show how the SGAN's accuracy is over 0.89 for almost all the considered time intervals.