Forecasting Volcanic Thermal States with LSTM and SGAN Using Multi-Source Satellite Time Series: The Vulcano Case Study (2016–2024)

Heat transfer at the surface in volcanic environments is an ongoing phenomenon representing the dynamic balance between the magma chamber and the adjacent rocks. In volcanoes, part of the magma’s energy drives fluid circulation, resulting in increased ground temperatures. Heat is primarily transferred through conduction, convection, and radiation, each detectable using specific techniques. Convection is evident in fumaroles and areas of diffuse degassing while moderate thermal anomalies indicate conductive heat transfer. Radiative fluxes can be measured using multispectral instruments. On Vulcano Island (Italy), the continuous monitoring network has recorded transient variations in heat flow from the active cone, associated with increased seismicity and ground deformation. Based on the generated time series, three volcanic thermal states have been defined (Background, Minor Crisis, and Unrest) corresponding to distinct thermal behaviors observed at the La Fossa crater. Building on these observations, we propose a two-stage methodology for forecasting volcanic thermal states using Artificial Intelligence applied to satellite remote sensing data. In the first stage, Long Short-Term Memory (LSTM) neural networks predict future values of time series derived from multi-sensor satellite imagery. In the second stage, a Semi-Supervised Generative Adversarial Network (SGAN), trained on the same satellite observations, classifies the LSTM-predicted series into volcanic thermal states. Input time series include established satellite-based monitoring products, such as the Normalized Thermal Index (NTI) and Volcanic Radiative Power (VRP) from VIIRS sensor, and environmental indices NDVI, NDWI, and NDMI from Sentinel 2 MSI sensor. This framework leverages the strengths of LSTM models for temporal forecasting and SGANs for robust classification with limited labeled data, enabling the prediction of volcanic thermal state evolution solely from Earth Observation data. Preliminary results indicate that the LSTM–SGAN framework can successfully forecast and classify thermal states at multiple future horizons. This work was supported by the 'Space It Up' project (code CUP I53D24000060005) funded by the Italian Space Agency, ASI, and the Ministry of University and Research, MUR, under contract n. 2024-5-E.0.