Computational Models for the Analysis of seismo-volcanic traces

This PhD thesis investigates computational models for the detection and picking of seismo-volcanic events from continuous waveforms. The research explores three complementary approaches: (1) a Grid-search method for optimizing Short Term Average over Long Term Average (STA/LTA) parameters, (2) an Active Learning (AL) strategy combining machine learning and human intervention to enhance classification accuracy, and (3) the application of a deep learning model, Earthquake Transformer (EQTransformer), based on transformer architectures, for local and regional seismic event detection. In the volcanic domain, the Grid-search method was applied to optimize the STA/LTA parameters for the detection of Explosion Quakes (EQs) at Stromboli volcano by using the overall measure proposed: Quality Numerosity index (QNi). The results demonstrated a substantial improvement in detection precision and consistency over standard parameter settings, with the optimized configuration increasing the QNi by more 0.24 on test data.The Active Learning methodology, using a Random Forest classifier, showed that human-guided relabeling and signal editing significantly improved the model’s classification performance, particularly in low-confidence events and hybrid signal scenarios. These enhancements enabled better discrimination between EQs and non-EQs and reduced the need for operator intervention over time. For tectonic seismicity, the EQTransformer model was tested both on local seismicity in the Groningen region (Netherlands) and on a regional dataset from the 2023 seismic sequence in Turkey. The model achieved high accuracy (up to 98\%) in event detection and phase picking, even on previously unseen data. Evaluation using Quality and Numerosity indices (Qi, Ni, QNi) confirmed that the deep learning model performed robustly across diverse seismic contexts, supporting its generalizability. Overall, this thesis demonstrates how the integration of traditional signal processing, machine learning, and deep learning can enhance the reliability of seismic and volcanic monitoring systems. The proposed methodologies are adaptable and scalable, offering valuable tools for real-time monitoring and early warning applications in both volcanic and seismic hazard assessment.