Unbalanced Learning: Deep Training Challenges in Seismic Phase Detection

In recent years, deep learning has become a popular technique in seismology, particularly for enhancing the automatic detection of seismic phases and building more complete and high-quality earthquake catalogs. Neural networks-based models, trained for P- and S-wave picking-, offer the potential to significantly improve the temporal and spatial resolution of seismic sequence (Tan et al., 2021; Beroza et al., 2021), lowering the magnitude of completeness and increasing the number of small earthquakes, providing complete seismic catalogues, useful and fundamental to seismic hazard analysis and fault geometry refinements.. The training of such models is a non-trivial process which requires a combination computational and geophysical skills, and some experience to understand how to tune the key hyperparameters to have a well-trained network. In this work, we investigate the training process behind such models for a specific case, highlighting the challenges and innovations involved in preparing deep learning algorithms for use in Italian seismological contexts. We present a series of experiments using the transfer learning technique, which combines the pre-trained model from the INSTANCE dataset (Michelini et al., 2021) with the new AQ2009 aftershock dataset (Bagagli et al., 2023), specifically developed from the L’Aquila 2009 sequence. The goal is to develop a more generalized model that is able to recognize multiple P- and S-wave arrivals typically occurring in Central Apennines seismic sequences A key challenge in training arises from the imbalance within AQ2009, where P-phase picks outnumber S-phase picks by a factor of four. This asymmetry introduces significant difficulty in achieving stable and generalizable models. To address this, we explore and present some strategies, with a specific focus on the Loss function computation, that seems promising in mitigating the effects of data imbalances, particularly when dealing with underrepresented classes in suboptimal training conditions. We report the most relevant training tests, examining the effects of the transfer learning method, class balancing, and the choice of hyperparameters on model performance. By analyzing how different training setups impact the network’s ability to recognize seismic phases, we aim to improve the quality of the models used to produce reliable seismic catalogues.