Using KLSH to Rapidly Search Large Seismic Signal Archives on a Desktop Computer

Abstract:

The use of waveform correlation detection has become increasingly important in the last decade, and as the basic calculation is straightforward, and the online archives of past signals are ever-increasing, the use of technique should only become more widespread. Yet there is an inherent limitation in how widely the method can be applied due to the computational demands of searching large signal archives quickly. In this study, we investigate the applicability of Kernelized Locality-Sensitive Hashing (KLSH) to significantly decrease the computational requirements to the point that searches can be done on a commodity desktop computer. KLSH probabilistically interrogates the database such that much of the database is ignored when searching for closest matches, thereby dramatically reducing the number of correlations that need to be calculated. We evaluate KLSH using data from the IMS primary station MKAR. First we built a KLSH indexed archive using all associated signals from the IDC LEB catalog for 2002-2013 (~308,000 signals). We then tested the signal matching capability using the ~26,000 IDC-detected signals from 2014, including a variety of regional and teleseismic phases (56% are teleseismic P). We used the LEB phase assignments as ground-truth to score the results. Using a simple 0.60 correlation threshold, requiring at least two archive matches, and applying screening criteria based on consistency of metadata of archive matches, we were able to robustly identify 12% of the 2014 signals, including many teleseismic P phases from a variety of locations. Comparing KLSH against a full search, we established a recall rate of > 0.9, with search time on the order of 10 ms.