Semi-supervised learning for proxy thermal-state classification on hydrothermal systems: The case of Vulcano (2016-2024)

Vulcano, an island in the Aeolian archipelago north of Sicily (Italy), has exhibited secondary volcanic activity since its last eruption (1888-1890), including the emission of noxious gases from fumarolic vents at La Fossa crater, Forgia Vecchia, and Levante Bay, as well as submarine hydrothermal manifestations near Levante Harbour. The dynamics of hydrothermal fluids periodically induced transitions in the volcanic system between background, minor crises and unrest phases. The ground-based monitoring network deployed on the island since 1990 has provided time series of fumarole temperatures, whose long-term interpretation has contributed to identifying the temporal evolution of different phases. In this context, spaceborne remote sensing has emerged as a reliable tool, providing the synoptic perspective and accuracy required to monitor the surface thermal evolution of complex hydrothermal systems such as Vulcano. This study presents an automated approach to classify thermal activity levels, categorized as Background, Minor Crisis, and Unrest, from satellite-derived thermal and environmental indicators as proxies for the system's hydrothermal state. A Semi-supervised Generative Adversarial Network (SGAN) is applied to different satellite-derived indicators, including the Normalized Thermal Index (NTI), and the Volcanic Radiative Power (VRP) from VIIRS sensor; the Normalized Difference Vegetation Index (NDVI), the Normalized Difference Water Index (NDWI), and the Normalized Difference Moisture Index (NDMI) from Sentinel-2 MSI sensor. Time-series of temperatures from the ground-based network allowed for labeling the different classes. Experimental results show that, from April 2016 to December 2024, the SGAN model achieved an overall Macro F1-Score of 97% in classifying thermal activity phases. This framework demonstrates significant potential for operational monitoring, providing a robust, space-based tool that complements traditional surveillance networks in detecting hydrothermal instability.