Eruptive dynamics of the 2024 lava fountains at Etna's Voragine Crater analyzed using unsupervised learning of multi-sensor satellite data

Lava fountains at the Voragine summit crater of Etna volcano in July–August 2024 were investigated using high-frequency satellite observations integrated with unsupervised machine learning techniques. Six paroxysmal events were analyzed using multisensor data from Sentinel-2 MSI, MSG SEVIRI, and Sentinel-5P TROPOMI, which enabled the retrieval of key eruptive parameters, including Volcanic Radiative Power (VRP), Time-Averaged Discharge Rate (TADR), erupted lava volume, volcanic cloud height, SO₂ mass emitted, and ash dispersion. Based on these eruptive parameters, an exploratory multivariate classification using the unsupervised 𝑘 -means algorithm is used as an objective and reproducible classification tool to identify representative eruptive patterns based on the joint behavior of the selected variables. Despite the small dataset inherent to the natural occurrence of the sequence, multivariate clustering served to objectively group the events into three representative behaviors: (1) “ paroxysmal unrest ”, characterized by long-lasting, low-explosivity events with substantial lava emission and extended pre-eruptive buildup; (2) “ transitional episodes ”, featuring moderate-duration episodes with intermediate lava volumes erupted, SO₂ emissions, and volcanic clouds heights; and (3) “ paroxysmal episodes ”, consisting of short-lived, highly explosive, SO₂-rich events producing high volcanic clouds and widespread ash. Comparison with volcanic glass composition supports a model in which explosivity is driven by variable influx of more primitive magma and by its mixing with the more evolved magma stored in the shallow reservoir, which in turn regulates the rest time between explosive episodes. These results provide new constraints on the evolution of paroxysmal activity at Voragine and offer a practical framework for classification of basaltic explosive activity using unsupervised learning methods on satellite data.