Cross-Regional Seismic Event Discrimination via Convolutional Neural Networks: Exploring Fine-Tuning and Ensemble Averaging

The gradual densification of seismic networks has facilitated the acquisition of large amounts of data. However, alongside natural tectonic earthquakes, seismic networks also record anthropogenic events such as quarry blasts or other induced events. Identifying and distinguishing these events from natural earthquakes requires experienced interpreters to ensure that seismological studies of natural phenomena are not compromised by anthropogenic events. Advanced artificial intelligence methods have already been deployed to tackle this problem. One of the applications includes Convolutional Neural Networks (CNN) to discriminate different kinds of events, such as natural earthquakes and quarry blasts. In this study, we investigate the effects of ensemble averaging and fine-tuning on seismic event discrimination accuracy to estimate the potential of these methods. We compare discrimination accuracy of two different CNN model architectures across three datasets. This was done with the best models from an ensemble of each model architecture, as well as with ensemble averaging and fine-tuning methods. Soft voting was used for the CNN ensemble predictions. For the transfer learning approach, the models were pretrained with data from two of the datasets (nontarget regions) and fine-tuned with data from the third one (target region). The results show that ensemble averaging and fine-tuning of CNN models leads to better generalization of the model predictions. For the region with the lowest numbers of one event type, the combination of ensemble averaging and fine-tuning led to an increase in discrimination accuracy of up to 4% at station level and up to 10% at event level. We also tested the impact of the amount of training data on the fine-tuning method, showing, that to create a global model, the selection of comprehensive training data is needed.