Microseismic monitoring : Hierarchical transdimensional tomography and propagation of velocity model uncertainties to seismic event locations

Abstract : Earthquake hypocentre locations are crucial in many domains of application (academic and industrial) as seismic event location maps are commonly used to delineate faults or fractures. The interpretation of these maps depends on location accuracy and on the reliability of the associated uncertainties. The largest contribution to location and uncertainty errors is due to the fact that the velocity model errors are usually not correctly taken into account. We here illustrate how probabilistic inversion approaches can lead to more reliable results in terms of seismic event locations, for a real dataset recorded in the context of hydraulic fracturing. In this work, reliable P and S velocity model uncertainties are first estimated with a hierarchical transdimensional Bayesian tomography of perforation shot data. We implement a sampling Monte Carlo type algorithm to generate layered velocity models distributed according to the posterior distribution. The number of layers is itself a model parameter of the inversion. The uncertainties on observed P and S traveltimes are considered as hyperparameters and are also inverted. The resulting velocity model uncertainties are compared to sonic logs that have been acquired in order to assess the quality of tomography results. In a second step, we propagate these velocity model uncertainties to the seismic event location in a probabilistic framework (Gesret et. al., 2015). This enables to obtain more reliable hypocentre locations as well as their associated uncertainties accounting for picking and velocity model uncertainties. Earthquake locations are more often computed using only picked arrival times as observed data. We here compare the resulting event locations when only arrival times are considered and when both arrival times and P-wave polarizations are considered. We use a recently proposed Bayesian formulation of the probabilistic event location that integrates properly the polarization information. This formulation relying on angular central Gaussian probabilities uses directly the covariance matrix (Gaucher et. al., 2016) rather than separating the azimuth and inclination angles. The uncertainty domains are thus much more reliable