“Failed” eruptions revealed by pattern classification analysis of gas emission and volcanic tremor data at Mt. Etna, Italy

Abstract

During the spring of 2007, paroxysmal activity occurred at the Southeast Crater of Mt. Etna, always associated with sharp rises in the amplitude of the volcanic tremor. Activity ranged from strong Strombolian explosions to lava fountains coupled with copious emission of lava flows and tephra. During inter-eruptive periods, recurrent seismic unrest episodes were observed in the form of temporary enhancements of the volcanic tremor amplitude, but they did not culminate in eruptive activity. Here, we present the results of an analysis of these inter-eruptive periods by integrating seismic volcanic tremor, in-soil radon, plume SO₂ flux, and thermal data. SO₂ flux and thermal radiation are envisaged as the “smoking gun,” and certifying that changes in seismic or radon data can be considered as volcanogenic. Short-term changes were investigated by pattern classification based on Kohonen maps and fuzzy clustering on volcanic tremor, radon, and ambient parameters (pressure and temperature). Our results unveil “failed” eruptions between February and April 2007 that are explained as ascending magma batches, which triggered repeated episodes of gas pulses and rock fracturing, but that failed to reach the surface.

Appendices

Appendix 1: Notes on in-soil radon measurements

Radon is a radioactive, noble gas emitted from soil and soluble in water. Especially in volcanic/geothermal areas with vigorous circulation of magmatic or hydrothermal fluids at shallow depth, in-soil radon emissions are the surface expression of convective flow of gases along fractures. Convection facilitates the transport of radon from depth to the surface, and the abundance of the radon transported depends also on the abundance of uranium-bearing rocks. In recent years, promising results were obtained by applying soil gas radon prospecting to the detection of hidden (but seismically active) and outcropping faults (Burton et al. 2004; Cigolini et al. 2007; Neri et al. 2007, 2011b; Giammanco et al. 2009; Siniscalchi et al. 2010). In active volcanic environments, increments in in-soil radon emission that are not related to environmental influences (mostly earth tides and/or changes in atmospheric parameters, such as rainfall, barometric pressure and air temperature) can be the expression of changes in the mass flux within hydrothermal systems (Durrance and Gregory 1990). These changes can be related to increased output of magmatic gas and/or to increased permeability of the rock due to newly formed stress-induced fractures. On Mt. Etna, continuous soil radon monitoring was also used to recognize precursory signals to paroxysmal eruptive events (e.g., Cox et al. 1980; Alparone et al. 2005; Cigolini et al. 2005; Neri et al. 2006; Giammanco et al. 2007).

Our station TdF was equipped with a Barasol™ probe (Algade, France) placed into a borehole at a depth of 1.6 m in pyroclastic deposits produced by the 2002–2003 eruption (e.g., Allard et al. 2006). According to the specifics given by the manufacturer, the borehole where the Barasol™ probe was installed was not isolated from the surface (the borehole is capped with a small semipermeable Mylar^® sheet in order to make gas flow out of the hole while preventing rain or snow from entering it). The probe is held some 40 cm above the borehole bottom, as in a fumarolic environment with a lot of water vapor, this setup lets the gas flow freely up along the borehole (and thence out of it through the semipermeable cap), preventing water vapor from condensing on the sensor. The probe at the TdF site was also modified by adding a robust filter close to the sensor, in order to prevent any condensation of water vapor on its surface.

Under steady flow conditions, water condensation would cause both an increase in soil temperature due to heat loss and an apparent increase in radon activity in the condensing zone, due to the lower proportion of residual water vapor and the higher proportion of non-condensable gases (particularly CO₂) after water condensation. However, as already proposed by Aubert and Baubron (1988), condensation of CO₂-rich hydrothermal fluids rising just south of TdF through a N–S-trending fault (namely, very close to the location of our radon probe) produces a soil temperature anomaly right on the fault plane and an eastward lateral flow of condensed water. This process, favored by the high soil permeability in the area, was traced at the surface by Aubert and Baubron (1988), measuring a concurrent and decreasing anomalous emission of CO₂ and radon along the presumed direction of water flow. Both CO₂ and radon were evidently released by the flowing water after being trapped in it, and hence were removed by water from the residual soil gas after condensation. If the latter mechanism is true, then the increased condensation of water vapor at shallow depth, coupled with increased soil temperature and decreased soil radon activity, would be caused by a decrease in the input of high-enthalpy magmatic fluids due to changes in volcanic activity. However, this would be in contrast with the overall increase in volcanic activity observed after the onset of the 2007 sequence of paroxysmal events. Therefore, we would be inclined toward the mechanism of radon dilution to explain the observed variations in temperature and radon activity. Further studies that will include monitoring of soil CO₂ will possibly clarify the interaction of temperature, condensation process, and radon activity.

Appendix 2: SOM and fuzzy clustering in KKAnalysis

Kohonen maps or SOM are unsupervised pattern classifiers, which “learn” from data of whatever nature provided in numerical format. SOM are made up of a number of nodes, each one representing a feature vector, and being prototypes of a number of patterns. During the so-called training phase, the node weights (feature vector elements) are adjusted minimizing the differences between the nodes and the original pattern. A pattern is assigned to the node for which the smallest difference is encountered; this node is referred to as the Best Matching Unit (BMU) for that pattern. Once KKAnalysis concludes the training phase of the SOM, the nodes are mapped into a 2D representation space applying the principal component analysis (Fig. 6a). KKAnalysis exploits an appealing property of SOM, often referred to as topological fidelity, which means that patterns represented by neighboring nodes in the SOM are close to each other also in the original data space. KKAnalysis applies a color coding to the nodes depending on their position on the map and assigns to each pattern the color of the BMU to which it belongs. Given the topological fidelity, similar patterns show similar colors. Accordingly, changes of pattern characteristics in time are visualized as output of KKAnalysis by creating a sequence of colored symbols.

Fig. 6

Principles of functioning of KKAnalysis: a Kohonen maps, b fuzzy clustering. The input data in a can be of whatever nature though expressed in a numerical form. In our application, data were given by the spectral content of volcanic tremor, radon values, temperature, and pressure. The vector components (red, green, and blue color) in b indicate to which degree, expressed by the results of fuzzy clustering (FC, ranging 0–1), a pattern belongs to various clusters (three in the example)

Along with SOM, KKAnalysis exploits the so called fuzzy clustering algorithm. This algorithm allows the potential allocation of a pattern into several clusters (or classes) according to a vector, which identifies its membership (Fig. 6b). A pattern with membership 1 belongs exclusively to one cluster. In case of membership <1, the vector components indicate to which degree a pattern belongs to various clusters. A membership <0.5 means rather fuzzy. Combined with the colors of the nodes of the SOM, fuzzy clustering allows the identification of transitional regimes between clusters, highlighting the development of the characteristics of a multidimensional feature vector.

Appendix 3: Application of KKAnalysis to seismic and radon data

Before using KKAnalysis, we preprocessed our seismic data recorded at ESPC (Fig. 1), calculating the spectral content of volcanic tremor from February 21 to April 20, 2007. The spectral analysis was carried out applying the Short Time Fourier Transform. We followed Langer et al. (2011), preparing feature vectors with 62 components, each of which corresponded to power spectral densities in frequency bins of ~0.29 Hz. The total frequency range covered was from 0.15 to 18 Hz. Each pattern represented 5 min of tremor data, with a total of 288 patterns per day. The whole amount of patterns obtained was 16,992, representative of 59 consecutive days. For the setting of KKAnalysis, we fixed the size of the SOM at 40 × 9 nodes, the same size chosen by Langer et al. (2011), and similarly we considered three clusters for the fuzzy clustering. Patterns representative for the 2007–2008 eruptive episodes analyzed by Langer et al. (2011) were considered as a reference data set, as they represented spectral characteristics of volcanic tremor shortly before, during, and immediately after climactic volcanic activity.

Exploiting our experience on volcanic tremor classification, we applied KKAnalysis to radon data along with barometric pressure and soil temperature measurements acquired at the same site with the same sampling rate, as the latter two parameters affect radon emission (e.g., Aubert and Baubron 1988). Also in this case, we preprocessed the data set for numerical purposes, applying a logistic normalization function to the raw data. Overall, we obtained some 16,000 patterns sampled with a time interval of 5 min, covering a total of 59 days.