Using seismic gradiometry and machine learning to discriminate between explosion and earthquake seismic sources

Abstract

Using seismic signals to discriminate between naturally occurring earthquakes and underground explosions has important applications in national security. Because of this, there is a rich literature on the detection and classification of seismic sources, almost all of which rely on conventional seismic networks and/or arrays and the three component, collocated seismograms recorded by them. For this work, we explore the efficacy of a using a different type of seismic data for the discrimination task. Specifically, we use the data recorded from a single seismic gradiometer as input to various machine-learning type classification algorithms. This is a different paradigm for the discrimination problem, as rather than relying on three-component ground motion collected at various points by a large-scale seismic network, we use the (up to) twenty unique components (e.g spatial gradients, dynamic strain, and changes in radiation pattern) derived from a single surface-deployed gradiometer, which occupies a footprint of only ~5% of a seismic wavelength. In a series of numerical tests using synthetic data, we use the gradiometric wave attributes as input data for off-the-shelf machine-learning algorithms to differentiate between buried explosions and earthquakes. We benchmark our gradiometric detection schemes to similar cases where the input data is collected from a conventional seismic array and find that the two data types return comparable discrimination performance when using well-known machine learning classification algorithms. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525