Machine learning-based seismic event classification at selected stations of the Czech Regional Seismic Network

The Czech Republic is a moderately active seismic region. Although most recorded earthquakes are weak, some events are strong enough to be felt by the population (e.g. Hlučín, December 2017, ML 3.5; West Bohemia, December 2025, ML 2.5 – 3.0). The majority of seismicity is mining-induced; however, areas of natural seismicity also exist, such as the Opava region and West Bohemia.

Seismic activity in the Czech Republic is monitored by several seismic networks, with the Czech Regional Seismic Network (CRSN) serving as the primary system. Seismic monitoring includes rigorous event classification, i.e. distinguishing between natural and induced seismicity as well as between earthquakes and surface explosions recorded by seismic stations.

With the growing volume of seismic data, semi-automated seismic event processing has become increasingly necessary. Automatic seismic event classification based on seismic signals represents a key step toward this goal. In previous work, we achieved promising results using machine learning (ML) techniques applied to data from the Ostrava-Krásné Pole station (OKC), which monitors the northeastern Czech Republic, an area with historically significant mining activity.

In this study, we extend seismic event classification to stations with a different instrumentation and apply newer ML approaches. Namely, we analyze data from the Moravský Beroun (MORC) and Vranov (VRAC) stations of the CRSN, both equipped with broadband STS-2 sensors with a lower corner period of 120 s and recording continuous seismic waveforms at 100 Hz. The studied dataset used for binary classification consists of records of mining-induced seismic events (8,338 from MORC, 4,085 from VRAC) and quarry blasts (4,193 from MORC, 3,041 from VRAC), which were localized in the Czech Republic and its neighbouring countries in 2023 – 2025. Induced events with known P- and S-wave arrivals and explosions with known P-wave arrivals were selected. The P-wave and S-wave arrival times were taken from bulletins provided by the Institute of Physics of the Earth.

Each processed event record includes 1 s before the P-wave arrival and either 20 s after the S-wave arrival (if available) or 30 s after the P-wave arrival. Data preprocessing included Z-score normalization and time-frequency transformation of the seismic signals.

We evaluated several models, including LSTM, LSTM-FCN, LSTM with an attention block, a hybrid CNN-Vision Transformer (CNN-ViT) neural networks, and XGBoost. The evaluated models achieved F1-scores of 0.92 (LSTM-based), 0.94 (XGBoost), and 0.96 (CNN-ViT), with comparable performance for MORC-only, VRAC-only, and combined datasets.

Furthermore, we combined data from the MORC and VRAC stations with records from the OKC station in a multimodal approach (37,561 events). Despite differences in instrumentation (e.g. lower corner periods of 120 s versus 30 s), the models achieved consistently high performance, with F1-scores ranging from 0.92 to 0.96 (CNN-ViT model yielding the best results).

These results demonstrate that machine learning models represent a promising step toward automated seismic event classification and more efficient seismic signal processing.