Explainable Seismic Event Discrimination: Improved Explainability With Vision Transformers

Abstract The discrimination of seismic events traditionally relies on manual expert analysis or cross‐correlation techniques for pattern recognition. Although flexible algorithms from deep learning, such as convolutional neural networks, have demonstrated high discrimination accuracy for different seismic event origins based on spectral characteristics, their use in automated discrimination remains limited. Two main challenges hinder further deployment: first, these models are typically not transferable across regions with differing event characteristics; second, deep learning models often lack interpretability due to their black‐box nature. In this study, we investigate vision transformers as a novel architecture for seismic event discrimination and compare their accuracy and explainability to a convolutional neural network baseline. Vision transformers leverage attention mechanisms offering a distinct approach to visualizing the decision‐making process of deep learning models. We evaluate both model types on seismic data from two regions in Germany: the Vogtland‐West Bohemian region where earthquakes along the Leipzig‐Regensburg fault zone are discriminated from quarry blasts (QBs), and the Ruhr Area, where mining‐induced events are discriminated from QBs. Our results show that vision transformers achieve high discrimination accuracy on test data from both regions (97%–98%) performing on par with the baseline. For explainability, we apply Attention Rollout and the gradient‐based LeGrad method for vision transformers and SmoothGrad for the baseline. Combined with occlusion techniques, these methods demonstrate that vision transformers analyze spectrograms more comprehensively. Our findings highlight the potential of vision transformers as both a high‐performing and interpretable alternative to convolutional neural networks for automated seismic event discrimination.