Statistical assessment of predictive modelling uncertainty: a geophysical case study

Abstract

When the results of geophysical models are compared with data, the uncertainties of the model are typically disregarded. This paper proposes a method for defining the uncertainty of a geophysical model based on a numerical procedure that estimates the empirical auto- and cross-covariances of model-estimated quantities. These empirical values are then fitted by proper covariance functions and used to compute the covariance matrix associated with the model predictions. The method is tested using a geophysical, spherical, thin-sheet finite element model of the Mediterranean region. Using a χ2 analysis, the model's estimated horizontal velocities are compared with the velocities estimated from permanent GPS stations while taking into account the model uncertainty through its covariance structure and the covariance of the GPS estimates. The results indicate that including the estimated model covariance in the testing procedure leads to lower observed χ2 values and might help a sharper identification of the best-fitting geophysical models.