Real-time classification of Irish seismic events using supervised machine learning

Seismic monitoring in areas of low natural seismicity is often complicated by anthropogenic signals which can have waveform characteristics similar to small magnitude earthquakes. In Ireland, quarry blasts make up the majority of signals detected by the Irish National Seismic Network (INSN), increasing analyst workload and making it difficult to produce reliable seismic catalogues.

Here we present an automated classification workflow for seismic events based on supervised machine learning methods that is suitable for operational use in discriminating between earthquakes, quarry blasts, and false detections. Instead of relying solely on waveform data, the classifier uses as input a combination of features derived from seismic waveforms plus event information, such as source parameters (e.g. depth, origin time, magnitude, number of phases) and location (e g. distance to the nearest known quarry).

Gradient-boosting classifiers, including XGBoost and CatBoost, were trained on a catalogue of more than 7,000 labelled events. Synthetic oversampling and hyperparameter optimisation were used to enhance robustness and reduce overfitting, and combining probabilistic outputs from multiple models was also explored. Our results show that by incorporating event information along with waveform features the accuracy and reliability of the classification was improved, with the final model achieving accuracies of more than 99% for quarry blasts and 95% for earthquakes, based on testing with unseen event data.

The trained classification tool has now been integrated into the INSN processing environment (SeisComP), and is currently being used to provide classification information for real-time alerts of automatically triggered events, as well as assisting in manual analysis. Although this work relies on the use of network-specific information, it demonstrates a transferable approach that can be used to integrate data-driven classifiers into operational geophysical monitoring systems. It also highlights the effectiveness of modern machine-learning techniques, supports the development of next-generation seismic data services and provides a practical example of how such tools can augment and complement traditional seismic monitoring workflows.