Search

In recent years, machine learning techniques have been widely applied in seismological data processing such as seismic event detection, phase picking, location, magnitude estimation, and further data analysis for determining source mechanisms. Especially in earthquake location, deep learning methods are used to reduce location errors compared to conventional algorithms.In this study, we present a deep learning based epicentral distance estimation with two separate models using seismic data from two stations as input data. The first model is the P- and S-wave arrival time picking model and the second is the epicentral distance estimation model. Since the traditional epicentral distance estimation methods uses the difference in arrival times between P- and S-waves, the P- and S-wave arrival times were first predicted from three-component seismic data so that this information could be directly used as the next input data. This picking information is used as input data along with the station location in the epicentral distance estimation model to output the final epicentral distance. Since this method uses data from two stations, it has higher accuracy than epicentral distance estimation using data from a single station.The P- and S-wave arrival time picking model was modified by referring to the ResUNet (Diakogiannis et al., 2020) structure to improve performance based on the seismic detection and phase picking model from the three-component acceleration data developed by Mousavi et al. (2020). This modified model performs feature extraction for P- and S-phase picking and includes a residual block and skip connection. The model for estimating the distance from the epicenter was constructed using a basic artificial neural network (ANN) architecture. As input data, a total of eight features were used by adding six combinations of the difference in arrival times of P-wave and S-wave in each component of the two stations and two values of the latitude and longitude difference between two stations. The ANN architecture consists of four hidden layers and the epicentral distances of the two stations are final output.The STEAD data were used as training data and test data. The STEAD is a seismogram dataset recorded from about 450,000 global earthquakes, and among them, data with magnitudes greater than 2.5 and epicentral distances less than 400 km were selected and used. As a result of applying the trained model to the test data, the mean absolute error of the predicted epicentral distance was 6.5 km, which showed improved performance compared to the previous results. Also, since this method uses six time-differences as input data, it can provide more robust results even in the presence of random noise at the picked times.

Seismic reflection images derived by ambient-noise seismic interferometry (SI) can show subsurface structures without active sources. To image and interpret the upper mantle structures and tectonic boundaries beneath the southern part of Korean Peninsula, we applied SI method to seismic ambient noise data recorded at 119 seismic stations on the Korean Peninsula in 2014 (from the seismic network of the Korean Meteorological Administration). The factor that makes interpretation difficult is the steeply dipping events in reflection images. Most of these events of apparent steeply dips show as true reflection events from steep geologic boundaries. Therefore, we need to attenuate these events to interpret true reflection events. These events overlap many times. Also, the value of the slope has several values close to half of the Rayleigh waves or P waves. To attenuate these events with these complex features, we used machine learning techniques. We attenuated our steeply dipping events by applying the Extraction of diffractions method. As the steeply dipping events are attenuated, horizontal events were strengthened, and noises were attenuated. We can more clearly identify the reflection events of the Moho discontinuity and the lithosphere/asthenosphere (LAB) boundary near the two-way reflection times of 7-11 s and 17-22 s respectively.

Earthquake Early Warning System (EEW) is a technology that calculates earthquake parameter using P-wave that arrives earlier and warns the expected damage area before the arrival of destructive S wave. Therefore, many countries are operating EEW to mitigate damage from earthquake shaking. Especially an on-site EEW is drawn attention as it can reduce blind zones due to using only a single or minimum station. In the on-site EEW, it is important to quickly predict the seismic intensity, which indicates the degree of ground damage, response of structures and ground shaking felt by people at a given location, rather than information on the magnitude or distance of the earthquake.In this study, we suggest a machine learning (ML) model that can directly estimate the seismic intensity scale from initial P-waveforms of three-component acceleration data measured at a single station. We used 1D-Convolutional Neural Networks (1D-CNN), which have been shown good performance in signal processing of speech and medical data which are similar to earthquake signals. K-Net and KiK-net datasets, recorded at stations in Japan, were used for training the ML model. Since the amount of data is enough and all of data are labeled with Japan Meteorological Agency Seismic Intensity Scale (IJMA), the datasets were used as training data in this study. The developed model produced fast and accurate results using only the three-component acceleration field data at a single station.In order to test applicability of the trained model to the new dataset acquired from other regions, the trained model was applied to the STEAD data which were recorded at stations distributed globally. When the trained model was applied to STEAD data directly, the prediction results were worse than those of K-Net and KiK-net data. The reason is that the characteristics of the ground and waveforms are different depending on the region. Therefore, to solve this problem, transfer learning was applied, and only the parameters of a fully connected layer of pretrained ML model were fine-tuned using small number of both labeled target dataset and training dataset used for pretraining. Moreover, by considering the imbalance problem of the training data for transfer learning, it was able to obtain better prediction results. Ultimately, this study shows the pretrained model with a specific region dataset can provide reasonable prediction of seismic intensity to new dataset acquired from other regions using transfer learning.