ISTerre seminar

Analysis of multidisciplinary signals with deep learning for the anticipation of eruptions at Piton de la Fournaise

Wednesday 20 November 2024 - 11h00
Matthieu Nougaret - 
--- 
Monitoring the activity of volcanic edifices is central to the mitigation of volcanic risks and hazards. Various methods, mostly relying on monitoring geophysical signals such as seismicity and ground deformation, are used on many volcanoes around the world. These monitoring methods generate large amounts of data. This represents a challenge for real-time analysis. For now, volcano observatories use automatic alerts, for instance triggered when seismic activity reaches a threshold such as at Piton de la Fournaise volcano, La Réunion island, France or at Colima volcano, Mexico. Multivariate data analysis may allow refining such monitoring tools and could be key to improve eruption forecasting and management. Machine learning may be a key method to perform multivariate analysis of time series recorded by volcano observatories. Its use could bring new insights and understanding and help better anticipating eruptions. Indeed, machine learning algorithms are particularly effective in analysing time series and detecting anomalies, such as it is done in other sectors to detect for instance bank frauds or network intrusions. During my thesis, I am testing if and how signals from seismicity, ground deformation and CO2 degassing can be combined with the help of deep learning to detect and forecast volcanic eruptions at Piton de la Fournaise. The idea behind is also to leverage possible unknow correlations between geophysical and geochemical signals. I have tested unsupervised anomaly detection as a proxy to detect possible pre-eruptive signals. Based on the results of this method, I have built a supervised neural network to achieve the classification of time series to detect precursory signals. Finally, I tested time series prediction to help at the anticipation and forecast eruptions. For this purpose, I trained different artificial neural networks to make predictions on the evolution of seismicity and deformation.

Organizing team : Ondes et structures

Salle Dolomieu, Maison des Géosciences, 38400 Saint Martin d'Hères