Advanced methods for real-time identification and determination of seismic events
MetadataShow full item record
Natural disasters pose an indistinguishable threat to populations all around the world, affecting ~200 hundred million every year, with earthquakes being the most deadly. Global seismic monitoring allows for robust real-time analysis to provide useful information about an event to assist in earthquake emergency response. Additionally it is an essential tool for monitoring anthropogenic seismic sources like nuclear weapons tests, the use of which can have disastrous effects on human life, ecological environments and public health, ramifications that can last for generations. The focus of this thesis is on characterizing and identifying unique seismic events in near-real-time using the waveforms of initial seismic phase arrivals from teleseismic stations, their derivative products like radiated earthquake energy and rupture duration, and machine learning (ML). This thesis is a compilation of several works addressing novel methods for seismic event identification of: global tsunamigenic earthquakes, uncharacteristically high-energy tsunami earthquakes, deep earthquakes, and underground nuclear explosions (UNE). First, I present the current Real-Time Earthquake Energy and Rupture Duration Determinations (RTerg) products and methodology applied to a case study of fast-rupturing tsunami earthquakes in the Solomon Islands, testing the robustness of the RTerg derivative waveform products and Tsunami Earthquake (TsE) discriminant threshold used for real-time analysis. Second, I show how peaks in RTerg energy flux curves from teleseismic stations and their differences in broadband and high frequency bandwidths can be associated with depth phase arrivals (P, pP, sP) to identify deep earthquakes, highlighting the potential for real-time depth determinations using first derivative waveform products without additional processing of waveforms. Next, I introduce nuclear explosion monitoring from a global network of stations, starting with the compilation of the first openly available and comprehensive UNE seismic waveform and event catalog termed GTUNE (Georgia Tech Underground Nuclear Explosions). GTUNE seismic records are sourced from declassified nuclear tests, previously published datasets and openly available waveforms and were assembled into a user‐friendly format compatible with most python‐based ML packages. The next contribution to this thesis is the development of a global UNE classifier using labeled P-wave seismograms from GTUNE. I trained a Convolutional Neural Network (CNN) to identify three classes: earthquake P-wave, nuclear P-wave, and noise. I found that the model can accurately characterize most events, finding over 90% of the signals in the validation set, even with limited training data. Lastly, I combine the thesis works described thus far and applied similar ML methodology to classify/predict deep earthquakes, using both a CNN and a Deep Neural Net (DNN), trained on both physical features of the energy flux time series (prominence and peak density) as well as the original waveforms. Results show better single station predictions using the original waveforms. By contrast, for full network determinations, the energy flux products perform the best, despite the smaller training dataset. We anticipate that ML models like our UNE and deep earthquake classifiers can have broad application for other “small data” seismic signals including volcanic and non-volcanic tremor, anomalous earthquakes, tsunami earthquakes, ice-quakes or landslide-quakes.