Forecasting volcanic eruption is perhaps one of the most challenging field in volcanology, not only because of the complex non-linear behavior and intrinsic unpredicitability of volcanoes (e.g. Sparks [2003]) but also because of our lack of direct observation on what is exactly happening underground. Yet, the need to provide accurate forecast is indeed fundamental especially for the civil protection agencies to mitigate risks and properly assess hasards.
In monitoring active volcanoes, one of the key parameters used in eruption forecasting is the magma overpressure. This can be traditionally inferred from the ground deformation data measured on the Earth's surface by applying a kinematic model and assuming a given crustal rheology. Observations alone cannot provide accurate information about the state of the system not only because they are embedded with noise but also because they can sometimes be sparse and partial due to the limitations in data acquisition. A numerical model is then necessary in order to interpolate information from the data. However, a model is often a simplified version of the complex reality therefore it is incorporated with errors. Moreover, volcanic systems are time-dependent. By applying kinematic models, we tend to lose huge amount of information that are often hidden within the time-series of a given dataset.
Data assimilation takes advantage of the complementary information provided by a dynamical model and the observations. It is a model-data fusion technique used to estimate the state of a system, sometimes given a priori information that are based on error statistics. Many data assimilation algorithms are now available, one of which is known as the Kalman Filter (KF) that is based on estimation theory. The use of non-linear models and the computational cost of the KF are often the limitation of its application to geophysical problems. Evensen [2003] developed the Ensemble Kalman Filter (EnKF) which then tackles the issues that are related to the classic KF. EnKF uses an ensemble of models to construct a Monte Carlo approximation of the mean and covariance of the variables and/or model parameters that one aims to estimate.
In this work, we assess the ability of EnKF using deformation data in forecasting volcanic unrest. As a first attempt, we focused on a specific dynamical model (e.g. two-chamber model, Reverso et al. [2014]) that well-describes the behaviour of frequently erupting basaltic volcanoes such as Grimsvotn volcano. We performed several synthetic tests to address the following: 1) track the magma pressure evolution at depth and estimate non-evolving as well as time-dependent uncertain model parameters, 2) properly assimilate GNSS and InSAR data, 3) the strengths and weaknesses of EnKF and how we can compare it with a Bayesian-based inversion technique (e.g. MCMC) and lastly, 4) the strategic way of combining EnKF and MCMC. We also present a first time application of data assimilation to real-case deformation data of Grimsvotn volcano in Iceland. Our results show that EnKF works well with the synthetic cases and there is a great potential in utilising the method for real-time monitoring of volcanic unrests.
| Type : | : | Séminaire invité |
| Thématiques | : | Interprétation géophysique |
| PDF version | : |  PDF version |
