Description
Traditionally, the discrimination of seismic signals has been carried out manually by expert analysts, who assign events – such as earthquakes, quarry blasts, and other anthropogenically induced events – to specific categories. This process often involves the use of cross-correlation techniques for pattern recognition. Although methods from the field of deep learning, particularly convolutional neural networks, have demonstrated significant potential in automating seismic event discrimination, their widespread application remains limited. Two main challenges hinder their further and frequent deployment: First, the difficulty of developing models that maintain high accuracy across different regional seismic networks. Second, the limited interpretability of such models, which often makes their internal decision-making processes non-transparent.
In this contribution, we investigate Vision Transformers as a novel deep learning approach for the discrimination of different types of seismic events. We present a comparative analysis between Convolutional Neural Networks and Vision Transformers with regard to their ability to distinguish between earthquakes, quarry blasts, and mining-related seismic events. Our results demonstrate that the tested vision transformer architectures can reach discrimination accuracies between 95% and 98%.
To enhance the interpretability of model outputs, we employ visualization methods that provide insight into the internal reasoning of the Vision Transformer models. Specifically, we apply Attention Rollout to track and aggregate attention scores across multiple transformer layers, and LeGrad, a gradient-based attribution technique that identifies input regions most relevant for the model’s predictions. While Attention Rollout directly utilizes the architecture’s inherent attention structure, LeGrad derives importance by analyzing the gradients of the output with respect to the input. Together, these approaches offer complementary perspectives on model behavior and contribute to a more comprehensible and transparent understanding of how seismic signals are discriminated by the deep learning models. This supports the potential of Vision Transformers as an interpretable and high-performing alternative to Convolutional Neural Networks in the context of automated seismic event discrimination.