Uncertainty Quantification in Tomographic Inversion of Near-Surface Seismic Refraction Data

Understanding the near-surface structure of the Earth requires accurate prediction of physical properties of the subsurface, such as velocity estimated from tomographic inversion of seismic refraction data. The predicted velocity values are often uncertain due to epistemic uncertainty in the inversion process (i.e., imperfectly known underlying physics) and aleatoric variability in the data (i.e., inherent noise in observations). Although seismic refraction is widely used in near-surface applications, the associated uncertainty is rarely quantified and presented alongside the inverted velocity tomograms. In this study, the effect of epistemic uncertainty due to local variability in the initial model and aleatoric variability due to first-arrival picking error on the velocity prediction uncertainty are investigated. A stochastic framework is implemented based on a statistical approach where multiple realizations of stochastically perturbed initial models and travel time picks are generated and the uncertainty in the predicted velocity models is quantified. The two sources of uncertainty are first studied independently and then the combined effect is investigated. The results show that both sources affect the posterior uncertainty, but the uncertainty in the initial model has a greater effect than picking error on the uncertainty of the posterior velocity model. In addition, joint analysis of both sources of uncertainty shows that the uncertainty in the inverted model depends on predicted velocity values, depths, velocity gradients and ray coverages.