Uncertainty-Aware Deep Learning Methods for Robust Discrimination Between Earthquakes and Explosions

Summary The reliable classification of seismic events is crucial to precise seismic cataloging and robust hazard evaluation. Recent advances in deep learning have achieved great success in seismic event identification, leveraging their exceptional ability to automatically extract and recognize features. However, existing deep learning approaches to seismic classification rely exclusively on deterministic models, which cannot quantify epistemic uncertainty, preventing the estimation of prediction confidence that is critical for reliability evaluation. To address this issue, in this paper, we develop two uncertainty-aware deep learning models for earthquake (EQ) vs. explosion (EP) classification using the DiTing 2.0 artificial intelligence training dataset: a Bayesian convolutional neural network (BCNN) and a dropout-based CNN (DropCNN). We also implement a conventional deterministic CNN as a baseline model for comparative analysis. The experimental results demonstrate that both the BCNN and DropCNN can achieve a classification accuracy comparable to the conventional CNN, while providing additional uncertainty metrics for estimation confidence of prediction. Crucially, their uncertainty scores increase markedly in terms of encountering misclassifications or out-of-distribution samples compared to correct classifications, enabling automatic rejection of unreliable predictions based on the uncertainty threshold setting, triggering human verification or alternative discrimination methods. We then apply the trained models to analyze suspicious explosion events in the DiTing 2.0 dataset. The BCNN and DropCNN results exhibits strong agreement, consistently identifying 79 EP and two EQ events and flagging the remaining samples as uncertain classifications needing further verification. Our findings demonstrate that deep learning methods incorporating uncertainty estimation not only maintain a high accuracy in seismic event discrimination but also provide uncertainty estimation. This capability significantly enhances the model’s reliability and decision-making value in practical applications.