Machine Learning for Seismic Signal Processing: Phase Classification on a Manifold

Abstract

In this research, we consider the supervised learning problem of seismic phase classification. In seismology, knowledge of the seismic activity arrival time and phase leads to epicenter localization and surface velocity estimates useful in developing seismic early warning systems and detecting man-made seismic events. Formally, the activity arrival time refers to the moment at which a seismic wave is first detected and the seismic phase classifies the physics of the wave. We propose a new perspective for the classification of seismic phases in three-channel seismic data collected within a network of regional recording stations. Our method extends current techniques and incorporates concepts from machine learning. Machine learning techniques attempt to leverage the concept of "learning'' the patterns associated with different types of data characteristics. In this case, the data characteristics are the seismic phases. This concept makes sense because the characteristics of the phase types are dictated by the physics of wave propagation. Thus by "learning'' a signature for each type of phase, we can apply classification algorithms to identify the phase of incoming data from a database of known phases observed over the recording network. Our method first uses a multi-scale feature extraction technique for clustering seismic data on low-dimensional manifolds. We then apply kernel ridge regression on each feature manifold for phase classification. In addition, we have designed an information theoretic measure used to merge regression scores across the multi-scale feature manifolds. Our approach complements current methods in seismic phase classification and brings to light machine learning techniques not yet fully examined in the context of seismology. We have applied our technique to a seismic data set from the Idaho, Montana, Wyoming, and Utah regions collected during 2005 and 2006. This data set contained compression wave and surface wave seismic phases. Through cross-validation, our method achieves a 74.6% average correct classification rate when compared to analyst classifications.