Search

In this paper, a scalable end-to-end tree boosting system called XGBoost has been applied for predicting the magnitude of an earthquake from the early part of earthquake waveform data. This model uses the features extracted from the early P wave phase of the records as an input. The model's effectiveness has been verified by using data on earthquakes occurring in the Eurasian plate of Japan Islands from 1996 to 2021. Feature engineering has given 29 new features identified from the early P wave phase of the record, which show a high correlation with the magnitude of an earthquake. The comparison of predicted and actual magnitude shows that a trained XGboost model, which uses a single input record for magnitude prediction, gives an average prediction error of 0.004 ± 0.57 for earthquakes in the test dataset. In contrast, the average prediction error of –1.1 ± 0.80 and –0.65 ± 0.69 has been obtained for the magnitude estimated from conventional τc and Pd methods using the same test dataset. It is further seen that the average predicted magnitude of a single earthquake of magnitude 4.5 and 6.1 (MJMA) obtained by using multiple nearfield records using XGBoost model is 4.58 ± 0.33 and 6.32 ± 0.29, which is close to the actual magnitude of the earthquake. The results presented in this paper clearly show that the structured data can be effectively used by complex machine learning or deep learning models to predict earthquake magnitude from single or multiple records. [ABSTRACT FROM AUTHOR]

In this paper, we explore the use of seismicity data on a single-station basis in site response characterization. We train a supervised deep-learning model, SeismAmp, to recognize and separate seismic site response with reference to seismological bedrock (VS  = 3.45 km s−1) in a broad frequency range (0.2–20 Hz) directly from single-station earthquake recordings (features) in Japan. Ground-truth data are homogeneously created using a classical multistation approach—generalized spectral inversion at a total number of 1725 sites. We demonstrate that site response can be reliably separated from single-station seismograms in an end-to-end approach. When SeismAmp is tested at new sites in both Japan (in-domain) and Europe (cross-domain), it achieves the lowest standard deviation among all tested single-station techniques. We also find that horizontal-to-vertical spectral ratio (HVSR) is not the optimal use of single-station recordings. The individual components of each record carry salient information on site response, especially at high frequencies. However, part of the information is lost in HVSR. SeismAmp could lead to improved site-specific earthquake hazard prediction in cases where recordings are available or can be collected at target sites. It is also a convenient tool to remove repeatable site effects from ground motions, which may benefit other applications, for example, improving the retrieval of seismic source parameters. Finally, SeismAmp is trained on data from Japan, future studies could explore transfer learning for practical applications in other regions. [ABSTRACT FROM AUTHOR]

Peak ground acceleration (PGA) is a key parameter used in earthquake early warning systems to measure the ground motion strength and initiate emergency protocols at major projects. The traditional P -wave peak displacement-dependent PGA prediction model (Pd-PGA model) tends to underestimate the PGA for large earthquakes because it cannot make full use of the fault continuity rupture information hidden in the time-varying process of ground motion. In this paper, a continuous PGA prediction long short-term memory (LSTM) neural network model is proposed. The model takes eight sequential features of stations that are proxies of the energy and other physical parameters as input and provides the recorded PGA at the station as the target output. A total of 5961 records from 119 earthquakes recorded by the Japanese Strong-Motion Earthquake Network (K-NET) in Japan are used to train the neural network and 3433 records from 73 earthquakes are used as the test set to verify the model's generalization ability. The results show that within the same data set, the residuals of the predicted PGA for the proposed model are smaller than those of the Pd-PGA model and that the problem of PGA underestimation is resolved. The prediction accuracy also improves with increasing sequence length, which indicates that the LSTM neural network learns the rules hidden in the time series. To further verify the model's generalization ability, the model performance is analyzed for an M 7.3 earthquake that was not included in the training or test data sets. The results show that the residuals of the predicted PGA for the event are consistent with those for the test data set, indicating that the model has good generalization ability. [ABSTRACT FROM AUTHOR]

Shikoku island, southwestern Japan lies in the western Nankai Trough and showcases along-strike segmentation of slow earthquake behavior. Whether the spatial variation of tremor behavior reflects the regional differences in structure/source properties and how much such differences can be recognized by the seismic signals themselves are two questions addressed in this paper. Taking advantage of advanced methods in recognizing and classifying signals using machine learning approaches, we attempt to answer them by conducting signal classification experiments in Shikoku. Based on the tremor catalog from 1 June 2014 to 31 March 2015, the tremors recorded in four different areas were treated as different classes and segmented into 60-s-long signals. The number of tremors in four different areas (A to D, from west to east) reached 15,000, 31,000, 10,000, and 16,000, respectively. To efficiently distinguish between tremors from different areas, we applied a k-nearest neighbor (k-NN) classifier with Fisher's class separability criteria to select the optimal feature subset. The resulting classification performance reached more than 90% at all 12 stations. We further designed a triangle test to select the features that can better represent the differences in source properties between areas. We found that the most efficient features were associated with (1) the number of peaks in the temporal evolution of discrete Fourier transforms and (2) the energy distribution in the autocorrelation function (ACF). To match the difference in behavior revealed by the ACF, the size of the tremor zone, which mainly controls how long the seismic energy lasts in a tremor episode, was determined to be largest in Area B and smallest in Area C. The heterogeneity of the asperities in a tremor zone, which may control how spiky the tremor signals developed over time, was determined to be strong in Areas B and C. Together with previously documented variations in slow earthquake behavior in the same area, we finally propose a conceptual model that provides a better understanding of the regional differences in the tremor source properties in Shikoku, Japan. [ABSTRACT FROM AUTHOR]