Machine Learning for Seismic Signal Processing: Phase classification of seismic events on a manifold

In this research, we considered the supervised learning problem of classifying seismic events according to seismic phase.  A seismic phase is a label given to a seismic event that characterizes the wave as either (1) a compression wave (e.g. P-wave), shear wave (e.g. S-wave), or surface wave (e.g. L-wave/R-wave) and (2) any reflections through the earth’s inner structure along the waves’ path.  In seismology, knowledge of the event arrival time and seismic phase leads to epicenter localization and surface velocity estimates useful in developing seismic early warning systems.

We propose a new perspective for the classification of seismic phases in three-channel seismic recordings.  Traditionally, methods have been developed where seismic features are estimated at multiple levels of resolution and used in statistical models for phase classification.  Our approach extends these ideas to incorporate concepts from machine learning.  Machine learning techniques leverage the concept of “learning” patterns associated with different data characteristics.  In this case, the data characteristics are the seismic phases. Our method used a multi-scale feature extraction technique for clustering seismic data on a low-dimensional manifold.  We then applied kernel ridge regression on the feature manifold for phase classification and designed an information theoretic approach for model parameter selection and boosting.

We applied our technique to a seismic data set from the Idaho, Montana, Wyoming, and Utah regions collected during 2005 and 2006.  This data contained compression wave and surface wave seismic phases.  Through cross-validation, our method achieved a 75% average correct classification rate when compared to analyst classifications.