Seismic Event Classification in North China Based on Machine Learning

Abstract

Automated seismic event classification is critical for earthquake monitoring, disaster mitigation, and global explosion detection. Distinguishing between tectonic earthquakes and explosions is particularly challenging in seismically active regions such as North China, as their waveform characteristics are highly similar. This study systematically compares traditional machine learning (ML) and deep learning (DL) approaches for seismic event classification in this complex environment. A comprehensive dataset was constructed, comprising 1,847 events and over 43,000 vertical component waveform records, including 1,022 tectonic earthquakes and 825 explosions. Training data from 2012 to 2016 were used to ensure a representative sample of seismic activity in North China. Feature extraction yielded a dataset of 40 statistical and spectral features. Simultaneously, time-frequency spectrograms were generated using the Generalized S-Transform (GST), constructing an image dataset of equal size for deep learning model training. The XGBoost model, applied to the feature dataset, achieved a classification accuracy of 95%, while EfficientNet, a convolutional neural network (CNN), reached 94% on the spectrogram dataset. Independent generalization testing on 2021–2022 data demonstrated that XGBoost retained high robustness, achieving 91% classification accuracy, outperforming EfficientNet’s 87%. Although both models exhibited strong classification capabilities, EfficientNet required significantly longer training time, whereas XGBoost involved moderate human intervention during feature extraction. Misclassification analysis showed minimal overlap, suggesting that the two models extract complementary features. Given these characteristics, integrating both approaches in future research could potentially enhance classification performance by leveraging their respective strengths. These findings provide a framework for applying advanced machine learning techniques to seismic event classification tasks in North China and similar regions worldwide. Future research should investigate ensemble models and interpretability techniques to further improve classification accuracy and model transparency.

Keywords: Seismic event classification, Machine learning, XGBoost, EfficientNet, Time-frequency analysis