Beyond traditional seismology : harnessing deep learning to analyze seismological data

Earthquake monitoring can be defined as the identification of earthquakes and estimation of parameters such as origin time, magnitude, source characteristics, etc. from continuous seismograms. The characterization of these parameters using only the first few seconds of P-wave data is particularly crucial for the purposes of earthquake early warning, which could alleviate the damages caused by hazardous earthquakes to life and property. Deep learning, with its ability to autonomously learn hierarchical representations from raw data, has delivered state-of-the-art results in fields of computer vision, speech recognition, natural language processing and many more. In a data-abundant field such as seismology, deep learning shows great potential in numerous research problems. This thesis covers a detailed analysis on the deployment and efficacy of deep learning based methodologies in rapid earthquake characterization and shear wave splitting analysis, through five research papers. The first paper explores a magnitude based classifier which categorizes seismograms as noise, low magnitude or high magnitude, with the decision boundary at local magnitude 5. In particular, the effect of using different lengths of data on the model's accuracy was investigated and no discernible dependence on the input data length was found. Several iterations of training were executed and accuracies upto almost 94% were reached. This finding was important to demonstrate that a combination of convolutional and recurrent layers is capable of extracting information about the earthquake magnitude from only the first few seconds of P-wave data and motivated us to develop the Convolutional Recurrent model for Earthquake Identification and Magnitude Estimation (CREIME) which is the topic of the next paper. CREIME is trained on 512 samples (5.12 s) of 3-component seismic waveforms which are either purely noise or contain 1-2 seconds of P-wave data preceded by pre-signal noise. The novelty of this method as a deep-learning based regressor for earthquake magnitude lies in the the use of a sequence as a data-label instead of a single value. The output of this model can be used to (a) discriminate between seismic signal and noise, (b) estimate the magnitude for detected events and (c) pick P-arrival times. CREIME is tested on two independent datasets and scores an average accuracy of 98% for event versus noise classification; it can estimate local magnitudes and first arrivals of P-waves with average root mean squared errors of 0.65 units and 0.13 s, respectively. It also outperforms the classical STA/LTA algorithm, typically used for earthquake detection, as well as previously published baseline models especially for the estimation of high magnitude. In the third paper, we introduce PolarCAP which is a convolutional autoencoder model focused on deciding the P-wave first motion polarity from the vertical component of waveforms. While this is a relatively simpler task that can be done easily by human experts, when it comes to large volumes of data, the manual processing becomes very time-consuming and tiresome while also giving way to human errors. We show that given a fixed window of 0.64 s duration, with the (known) P-wave arrival at its center, the deep learning model can learn to automatically and efficiently classify the first motion polarity as `positive' (up) or `negative' (down) with a 98% agreement with human assigned labels. This not only speeds up the process greatly, but also turns out to be more reliable than manually assigned labels as demonstrated by the analysis of almost 600 traces where the model predictions and published data labels are in disagreement. We find the model to be correct in ∼41% cases as opposed to only ∼28% cases where the assigned labels are correct. In the remaining 31% cases the polarity was ambiguous due to emergent arrivals being masked by background noise. PolarCAP leverages the vast Italian Seismic Dataset (INSTANCE) to learn the polarity mapping in a supervised fashion and achieves better recall than a previous study for both positive and negative polarities on a dataset from South California Seismic Network (SCSN), demonstrating its generalizability. The concepts explored in the first three papers form the building blocks for an automatic seismic data processing pipeline when integrated with DynaPicker, a model designed to pick P- and S- arrival times from continuous seismic waveforms. This pipeline is embedded within an open-source python package called SAIPy which is presented in the fourth paper. Aside from having functionalities to apply the entire pipeline or individual models to continuous waveforms, SAIPy also provides an user-friendly interface to use benchmark seismological datasets such as the STanford EArthquake Dataset (STEAD) and the Italian seismic dataset for machine learning (INSTANCE). We demonstrate the application of this pipeline to the aftershock sequences of the 2023 Turkey earthquake recorded at a station in Ceyhan, Turkey and observe that SAIPy not only successfully detects and satisfactorily estimates the magnitude of events which have been recorded in the catalogues, but also identifies uncatalogued events. Finally, the fifth paper explores another sub-field of seismology - shear wave splitting (SWS) analysis - which has been relatively uncharted by deep learning. We introduce a baseline deep learning model called SWSNet which is trained on synthetic data with the aim of determining the splitting parameters, fast-axis orientation (φ) and delay-time (δt), from pre-selected waveform windows. To ensure the generalizability of such a model to real data we employ a novel deconvolution/reconvolution method to minimize source and path effects from the real data and only retain the splitting effects which, in principle, makes it resemble the synthetic theoretical waveforms that the model was trained on. We compare the performance of this method in calculating station-averaged splitting parameters calculated from teleseismic waveforms recorded at the the USArray Transportable Array with previously published results and find the mean absolute differences in φ and δt to be 9.6° and 0.16s, respectively.