Searching for Unknown Earthquakes in the Guy-Greenbrier, Arkansas, Earthquake Sequence using Efficient Waveform Similarity Search

Recent seismicity rate increases in the central United States have been attributed to injection of wastewater from oil and gas production. One example is the Guy-Greenbrier, Arkansas, earthquake sequence, which occurred from July 2010 through October 2011, and was potentially induced by fluid injection into nearby disposal wells (Horton, 2012). Although the Arkansas seismic network is sparse, a single 3-component station WHAR recorded continuous data before, during, and after this earthquake sequence at distances ranging from 2-9 km. Huang and Beroza (2015) used template matching to detect over 100 times the number of cataloged earthquakes by cross-correlating the continuous data with waveform templates based on known earthquakes to search for additional low-magnitude events. Because known waveform templates do not necessarily fully represent all seismic signals in the continuous data, small earthquakes from unknown sources could have escaped detection. We use a method called Fingerprint And Similarity Thresholding (FAST) to detect additional low-magnitude earthquakes that were missed by template matching. FAST enables fast, scalable search for earthquakes with similar waveforms without making prior assumptions about the seismic signal of interest. FAST, based on a data mining technique, first creates compact "fingerprints" of waveforms by extracting discriminative features, then uses locality-sensitive hashing to organize and efficiently search for similar fingerprints (and therefore similar earthquake waveforms) in a probabilistic manner. With FAST, each search query is processed in near-constant time, independent of the dataset size; this computational efficiency is gained at the expense of an approximate, rather than exact, search. During one week of continuous data at station WHAR, from 2010-07-01 to 2010-07-08, FAST detected over 200 uncataloged earthquakes that were not found through template matching. These newly detected earthquakes have the potential to add insight into how earthquake sequences begin, and can provide sensitive, high temporal resolution observations of seismicity evolution with sparse instrumental coverage.