Data Mining in the Context of Monitoring Mt Etna, Italy

Vienna, Austria

The persistent volcanic activity of Mt Etna makes the continuous monitoring of multidisciplinary data a first-class
issue. Indeed, the monitoring systems rapidly accumulate huge quantity of data, arising specific problems of andling
and interpretation. In order to respond to these problems, the INGV staff has
developed a number of software tools for data mining. These tools have the scope of identifying structures in the
data that can be related to volcanic activity, furnishing criteria for the identification of precursory scenarios. In
particular, we use methods of clustering and classification in which data are divided into groups according to apriori-
defined measures of similarity or distance. Data groups may assume various shapes, such as convex clouds
or complex concave bodies.The “KKAnalysis” software package is a basket of clustering methods. Currently, it is
one of the key techniques of the tremor-based automatic alarm systems of INGV Osservatorio Etneo. It exploits
both Self-Organizing Maps and Fuzzy Clustering. Beside seismic data, the software has been applied to the geochemical
composition of eruptive products as well as a combined analysis of gas-emission (radon) and seismic
data.
The “DBSCAN” package exploits a concept based on density-based clustering. This method allows discovering
clusters with arbitrary shape. Clusters are defined as dense regions of objects in the data space separated by regions
of low density. In DBSCAN a cluster grows as long as the density within a group of objects exceeds some
threshold. In the context of volcano monitoring, the method is particularly promising in the recognition of ash particles
as they have a rather irregular shape. The “MOTIF” software allows us to identify typical waveforms in time
series, outperforming methods like cross-correlation that entail a high computational effort. MOTIF can recognize
the non-imilarity of two patterns on a small number of data points without going through the whole length of data
vectors.
All the developments aforementioned come along with modules for feature extraction and post-processing. Specific
attention is devoted to the obustness of the feature extraction to avoid misinterpretations due to the presence
of disturbances from environmental noise or other undesired signals originating from the source, which are not
relevant for the purpose of volcano surveillance.