Implicitly localized MCMC sampler to cope with nonlocal/nonlinear data constraints in large-size inverse problems

Emmanuel Cosme (UGA-CNRS-IRD-GINP/IGE, France)


Jean-Michel Brankart (UGA-CNRS-IRD-GINP/IGE, France); Pierre Brasseur (UGA-CNRS-IRD-GINP/IGE, France); Sammy Metref (UGA-CNRS-IRD-GINP/IGE, France)

Event: 2020 Ocean Surface Topography Science Team Meeting (virtual)

Session: Quantifying Errors and Uncertainties in Altimetry data

Presentation type: Type Forum

In high-dimensional inversion problems, sample-based Bayesian inversion methods are limited by the sample size. Ensemble assimilation of geophysical data is generally implemented with covariance localization techniques, which limit the geographical extent of the impact of each individual observation in the analysis. This is actually acceptable with local observations, but not with non-local observations. This work investigates a localized Monte Carlo Markov Chain (MCMC) sampling method that unifies the notions of covariance localization and non-local observations. First experiments to reconstruct Sea Surface Height from satellite altimetry are performed.
Emmanuel Cosme