Geophysics seminar: Nader Shakibay Senobari (UC Riverside) "Advances in earthquake similarity search: mining large seismic data sets from central and northern California"


 
01/15/2019 - 12:00pm
Location: 
Munk Conference Room
Event Description: 

Talk Abstract:

My research focuses on applying new similarity search methods and techniques for better
understanding earthquake and fault behavior in central and northern California.


In the first part of my talk, I will discuss my search for repeating earthquakes (REs) using
similarity search on recorded seismic waveforms from the northern San Francisco Bay Area.
Evidence from the San Andreas fault and elsewhere indicates that REs are correlated with, and
likely driven by, aseismic slip (creep) at depth. Source information of REs can therefore be used
as constraints on interseismic slip and seismic hazard models. By using a new fast similarity
search algorithm developed specifically for probing big seismic data sets, we found 198 RE
groups. Our method is successful despite the somewhat sparse seismic network in this region
as we detect and cluster RE pairs at different stations, rather than requiring that all events of an
RE sequence be detected on a few common stations. Our results can not only help us to map
the depth and extent of creep on several major faults but also reveal previously unknown
structural complexity – e.g. that subparallel strands of the Maacama fault are active and creep
simultaneously. We also recalibrated the empirical relationship between creep rate and RE
magnitude and recurrence interval using a shallow RE family and an estimate of the shallow
creep rate from InSAR data. Our creep rates estimated by this method are broadly in agreement
with creep rates reported in studies based on geodetic data, but add information on the creep
distribution at depth.


In a second, related similarity search project, I will show an alternative, efficient method for
searching for templates for matched filtering – the similarity Matrix Profile (MP). The MP is
essentially a report of the index (position) and correlation coefficient (CC) value of the most
similar portion within a time series for every subwindow it contains. Being optimized for GPU
clusters, the MP can be computed for very long time series (hundreds of days of data) in a
matter of hours, and can detect both low- and high- signal-to-noise events, as the CC values for
background noise tend to be low compared to earthquakes. Therefore, the MP approach, with
an appropriate CC threshold, can be used to detect earthquake swarms, aftershocks, and
foreshocks, low-frequency earthquakes (LFEs), repeating events and more. We use our new
approach to revisit the behavior of the 2004 Mw 6.0 Parkfield earthquake foreshocks and
aftershocks; we detect ~16 times more events than are listed in the catalog. I will also discuss
how our new approach can be used for creating an alternative LFE catalog for this area.

For more information on this event, contact: 
Peter Sharer
Event Calendar: 
sharknado