Extraction of body wave arrivals from seismic interferometry using time series features and machine learning methods and their applications to seismic tomography.


The development of seismic interferometry methods in the last decade has increased the utility of passive seismic surveys for seismic modeling and monitoring. Since passive seismic surveys do not require manual efforts in the field after receivers are deployed, they have the potential to be automated and inexpensive. Body waves are typically used in high-resolution surveys, as opposed to surface waves, but the extraction of body wave arrivals from ambient seismic noise is challenging. Studies that successfully extracted body wave arrivals from passive seismic data are few, and their processing often requires significant manual effort and/or prior information about subsurface velocities. In our study, we develop automatable methods based on machine learning that identify important time series features to extract body wave arrivals. In addition, we show an application in which we use body wave arrivals extracted via seismic interferometry to model the subsurface. Our first study focuses on developing a P velocity model of the Texas Gulf Coast passive margin using body wave arrivals from seismic interferometry, along with a variety of arrivals that add complementary constraints to different parts of the model. Our use of different data types overcomes challenges posed by the thick layer of sediments on the Texas Gulf Coast, which attenuates seismic energy. We discover a high-velocity body in the crust collocated with the Houston magnetic anomaly and speculate that this body formed during the Ouachita orogeny. In our second study, we find causal and acausal time series features that identify cross-correlated noise panels in which body wave energy dominates (over surface wave energy and other types of noise) from continuous seismic data acquired at the San Emidio, Nevada geothermal field. Using just 25% of the data we were able to produce higher quality virtual source gathers, with better signal-to-noise characteristics and more identifiable body wave arrivals, than could be produced using all recorded data. The highly automatable nature of these methods makes them easy to implement in real-time seismic data processing. In our third study, we compute three types of spatio-temporal features, or characteristics, of data recorded by the dense, nodal “Sweetwater Array” deployed in Texas in 2014. We use these features to carry out unsupervised clustering using k-means to separate clusters with dominant body wave energy. Subsequently, we extract the body wave arrivals to produce a virtual source within the array using seismic interferometry.



Seismic interferometry. Machine learning. Time-series. Seismic tomography. Ambient seismic noise. Geophysics. Inversion.