Analysis of multivariate time series at Piton de la Fournaise using machine learning

French active volcanic structures are closely monitored by the volcanological and seismological observatories of the Institut de physique du globe de Paris (IPGP). Various geophysical methods (for example, seismology and deformation) and geochemical methods (such as gas emission monitoring) constitute indicators of volcanic activity and allow for anticipating eruptions. The interpretation of this data guides authorities in volcanic crisis management. However, human analysis is becoming increasingly difficult due to the growing quantity and diversity of data, as well as the complexity of linking observations of different natures.

Within the framework of this thesis, we exploited surveillance data from Piton de la Fournaise, on La Réunion Island, in order to evaluate the possibility of detecting, or even predicting, the occurrence of volcanic eruptions. I developed and tested different machine learning algorithms applied to seismic time series (daily number of volcano-tectonic earthquakes), deformation (GNSS baseline measurements) and geochemical (CO2 flux in soil) data. Several approaches were explored : classification, anomaly detection and time series prediction. The results highlight both the potential of machine learning for studying these time series and the difficulties related to their highly volatile nature. Among the different methods, supervised classification appears as the most promising, although it still requires developments to be fully integrated into operational volcanological surveillance.