Archives: Presentations

We present a 2D travel time tomography method which models rectangular groups of slowness pixels from discrete slowness maps, called patches, as sparse linear combinations of atoms from a dictionary. We further propose to adapt the dictionary atoms to slowness data using dictionary learning, a form of unsupervised machine learning. We call this sparse regularization the local model, as it constrains the local or small-scale relationships of the pixels. Relative to conventional tomography methods, which constrain models to be exclusively smooth or discontinuous, the sparse local model permits smooth and discontinuous features where warranted by data. In this locally-sparse travel time tomography (LST) method, the local model is integrated via an averaging procedure with the overall slowness map, called the global model, which constrains large-scale features using L2-norm regularization. Since LST does not enforce global pixel correlations, the global slowness is obtained by inverting a sparse tomography matrix, which requires less computation than dense matrices of equivalent dimension. With efficient dictionary learning algorithms, the LST method scales well to tomography problems with large numbers of rays and pixels. We develop a maximum a posteriori formulation for LST, which is solved as an iterative inversion algorithm. We apply the LST approach to ambient noise tomography on a dense seismic array located in Long Beach, California (2011) and obtain high resolution Rayleigh surface wave phase speed maps of the region. This ‘large N’ array contained more than 5200 stations resulting in ~14 million unique ray paths. The new maps obtained are consistent with previous results, and show increased contrast along fault lines. Further, the distribution of phase speeds show strong correlation with underlying oil fields in the area (Long Beach, Wilmington), indicating that the Rayleigh waves resolve surface sediments related to the hydrocarbon areas.

Slidecast:

Earthquake detection remains one of the fundamental operations in observational seismology. Despite the explosion in quality and quantity of seismic data our ability to build dense and complete seismicity catalogs remains limited. Recent approaches exploit the self-similarity of earthquakes and rely on using the waveforms of a few well known events as templates. Terabytes of continuous seismic signals at multiple stations are cross-correlated against these templates and thousands of events with high waveform similarity may be easily identified. This procedure, known as template matching, is computationally expensive and fails for events whose waveforms are significantly different from the templates. Also, the magnitudes of completeness of intermediate-depth earthquake catalogs are typically large. A dense catalog with a small magnitude of completeness not only allows for a detailed study of seismicity patterns, but also enables the mapping of small-scale b-value anomalies. We propose a Convolutional Neural Network (CNN) architecture for the simultaneous detection of intermediate-depth earthquakes, the precise picking of p- and s- wave arrival times and the estimation of their local magnitude. We test our implementation using a synthetically generated dataset and propose a training scheme that leverages both real and synthetic data for optimal results. As an example we apply our technique to a cluster of intermediate-depth intraplate earthquakes in northern Chile. We are able to detect eight times more events than in the initial catalog and use the picked arrival times to relocate them. We clearly resolve a double-planed structure at depth. Furthermore, we build a detailed b-value map and find that high b-value anomalies correlate well with regions were dehydration is expected.

Slidecast: