A new study from the National Institute of Geophysics and Volcanology (INGV) uses an integrated approach to recognize signs of changes in the island's hydrothermal activity.
A new study dedicated to the island of Vulcano introduces an approach that, thanks to artificial intelligence and the integration of satellite data with measurements acquired on the ground, can allow for improve monitoring of the hydrothermal system, that is, the combination of water, steam and gas present underground.
The research, coordinated byNational Institute of Geophysics and Volcanology (INGV) in collaboration with the Department of Mathematics and Computer Science, University of Catania, was created within the framework of the SAFARI project (An Artificial Intelligence-based StrAtegy For volcAno hazard monItoring from space) funded by the INGV Dynamic Planet program, and was published in the scientific journal Remote Sensing Applications: Society and Environment.
The study analyzed data collected between 2016 and 2024, combining information on temperature and environmental conditions derived from the VIIRS and Sentinel-2 satellites with fumarole temperatures recorded by the INGV monitoring network in the La Fossa Crater area. explains Francesco Spina, INGV researcher and corresponding author of the research.
The use of a semi-supervised learning model allowed us to accurately distinguish the different activity conditions of the hydrothermal system: background,, minor crisis and unrest.
“In particular – he continues Gaetana Ganci, INGV researcher and co-author of the study – The use of a semi-supervised model based on generative neural networks (SGANs) allowed us to overcome the limited availability of labeled data, due to the rarity of crisis phases. The model, in fact, can learn effectively both with a small amount of labeled data and with a large amount of unlabeled data.
Le generative neural networks (SGAN)In fact, They are systems capable of recognizing different situations even with a few already classified examples, exploiting the information contained in unlabeled data.
The results show how Artificial intelligence applied to satellite data could support volcano monitoring, allowing us to analyze surface temperature variations over time and identify changes related to the activity of the hydrothermal system, paving the way for more advanced surveillance systems and the early identification of signs of instability.
Useful links:
National Institute of Geophysics and Volcanology (INGV)
Department of Mathematics and Computer Science, University of Catania
Detailed architecture and operational flowchart of the Semi-Supervised Generative Adversarial Network (SGAN). The diagram illustrates the two-pronged training process: (i) the Unsupervised Mode, in which the Discriminator/Classifier distinguishes between real samples (unlabeled data) and fake samples synthesized by the Generator from a latent vector z; and (ii) the Supervised Mode, in which labeled data is used to train the model to classify inputs into n specific classes. The feedback loop, labeled 'Fine Tuning' in the diagram, represents the iterative process of adversarial training and weight updating via backpropagation.
ASynoptic images of Vulcano island dated 26/11/2016 00:48 (top left), 03/02/2021 01:00 (top right), 04/10/2021 00:54 (bottom left), and 10/12/2021 00:48 (bottom right). The color scale indicates the NTI value of individual pixels. The 16 pixels within the green box are part of the Thermal Summit Area (TAS) of the active cone, and their NTI values were included in the dataset for training and testing the SGAN model. BBar chart of the classification results obtained from the trained SGAN model. Green bars represent time intervals classified as Background, yellow bars indicate those classified as Minor Crisis, and red bars correspond to time intervals classified as Unrest. Black boxes highlight periods identified as crisis phases (i.e., classified as Minor Crisis and/or Unrest), which include: (a) November 9, 2016–January 3, 2017; (b) March 6, 2018–August 28, 2018; (c) May 25, 2019–July 15, 2019; (d) July 8, 2020–August 28, 2020; (e) May 10, 2021–September 13, 2023; (f) May 4, 2024–June 27, 2024.