The Hydrothermal System of Solfatara Crater (Campi Flegrei, Italy) Inferred From Machine Learning Algorithms

Two machine learning algorithms were applied to three multivariate datasets acquired at Solfatara volcano. Our aim was to find an unbiased and coherent synthesis among the large amount of data acquired within the crater and along two orthogonal vertical NNE- and WNW-trending cross-sections. The first algorithm includes a new approach for a soft K-means clustering based on the use of the silhouette index to control the color palette of the clusters. The second algorithm which uses the self-organizing maps incorporates an alternative method for choosing the number of nodes of the neural network which aims to avoid the need for downstream clustering of the results of the classification. Both methods achieved an objective characterization of the shallow hydrothermal system of the volcano, enhancing and highlighting subtle geophysical anomalies likely correlated to structural pathways of deep magmatic degassing. Comparison between the results of K-means and self-organizing maps on the datasets with the largest number of nodes confirms that, with respect to the K-means, self-organizing maps compress the data in a way that better highlights finer details of the original data. However, the choice of the coloring scheme of the neurons is critical for an effective visualization of the results. Unsupervised integration of the three multivariate datasets allowed us to spatially correlate, with a high-degree of confidence, the geophysical anomalies recorded at the surface of the crater with those recorded at the subsurface along the two cross-sections. it also allowed us to associate those anomalies to different hydrothermal features such as shallow gas-saturated and water-saturated zones and their underlying fractures/faults feeding system. Our results suggest that the main shallow structural patterns, which influence the hydrothermal dynamics at Solfatara volcano, remained substantially unchanged in the last 13 years. Our approach shows that the use of clustering methods to interpret multivariate data reduces interpretation uncertainties and achieves an improved understanding of the complex dynamics occurring in volcanoes.