Signal-based Bayesian Seismic Monitoring

Perception can often be framed as the inverse of a better-understood forward process. For example, computer vision can be viewed as the inverse of computer graphics: a rendering engine computes a map from scene descriptions to image pixels, and vision systems attempt to invert this map. Bayesian inference provides a powerful framework for attacking inverse problems by exploiting the structure of the corresponding forward process, as represented by a generative probability model.

In this talk I'll describe a Bayesian approach to detecting and locating seismic events directly from seismic waveforms. We define a generative model consisting of a prior distribution over the location and attributes of seismic events, along with a probabilistic "rendering model" that specifies a distribution over signals observed by a geographically distributed network of stations, given a set of hypothesized events. Our model combines physics-based knowledge, such as the predictable travel times of multiple seismic phases, with statistical components, e.g., using Gaussian processes to represent correlated waveforms from nearby events. Inverting the model via a combination of Bayesian inference techniques (message-passing and reversible jump MCMC) allows us to detect and locate weak events from noisy data, even from only a single station. In addition to scientific applications, this capability is of particular interest in monitoring underground nuclear test bans such as the Comprehensive Test Ban Treaty.

Speaker: Dave Moore, UC Berkeley