Neural networks-based ‘first-response’ detection of volcano-tectonic seismic events in poorly monitored active volcanoes: the case of Pico de Orizaba (Citlaltépetl), Mexico

Abstract

Volcano-tectonic events (VTs) detection is crucial for real-time volcano monitoring, yet many relatively quiet volcanoes remain poorly instrumented, limiting the detection effectiveness. Pico de Orizaba (Mexico) exemplifies such a challenge, as its sparse seismic network has hindered comprehensive monitoring efforts for three decades. This study applies artificial and convolutional neural networks (ANN and CNN) to improve VTs detection at Pico, demonstrating how ANN/CNN-based tools can compensate for limited instrumental coverage and human resources. Using Pico's six-years (2019–2024) seismic waveforms datasets (raw, 5–15 Hz band-pass filtered, and spectrograms), powered by one-year data (2018) from the neighboring erupting Popocatépetl volcano, we trained and validated the ANN/CNN to classify VTs. The results indicate that both models successfully identified subtle waveform features indicative of VTs. The CNN-based classifiers offered 95–99 % accuracy, 97–100 % recall, and 93–98 % specificity with a F1-score of 0.95–0.99 in a setting where VTs are often obscured by high noise levels and data sparsity. Moreover, our results have proved possible to employ data from an erupting volcano to accurately (92–95 %) identify VTs from a contrastingly quiet volcano. Future efforts will focus on expanding training datasets, and refining model adaptability across different volcanic settings and multiple seismic sources. The deployment of ANN/CNN techniques at Pico represents a significant step toward improving VTs detection in under-monitored volcanic regions, ultimately contributing to more effective hazard assessment strategies.