Classification Strategies Applied to Event Nucleation and Phase Association

Machine learning is an emerging tool in seismology that can lead to better automatic decision making in seismic processing algorithms. Here we assess the applicability of machine learning to real-time seismic systems. While there is an opportunity for machine learning to be applied to many facets of a real-time system this abstract focuses solely on event location. In a real-time setting locating an earthquake requires three phases. The first phase, nucleation, will take potentially disparate arrival time estimates and attempt to seed event location(s). With earthquake location(s) proposed the second phase, association, matches picks to hypocenters and, by extension, assigns phase labels. With seismic phases labeled and tied to corresponding events, the third analysis phase, traditional event location, can be performed. At each step in the workflow the processing algorithm must make choices that amount to classification; e.g., declaring a seeded event as real or spurious, classifying a pick as a P or an S phase, determining whether a phase arrival belongs to event A or event B. While these questions are traditionally answered with simple thresholding strategies, it remains true that these questions naturally fit into the framework of machine learning. Consequently, in a local monitoring setting, we investigate the suitability of a naive Bayes image classifier for determining the number of potentially nucleated events, the effectiveness of logistic regression for distinguishing between P and S waves from travel times, and the applicability of discriminant analysis for determining to which event a phase arrival belongs.