Robust classification of blasts, collapses, and natural earthquakes via Siamese neural network

Summary

Monitoring the activity of non-natural seismic events is crucial for constructing an accurate seismic catalog, overseeing the safety of industrial operations, and mitigating the potential threats to local residents. However, recent studies have shown that neural network models trained on local data sets may not generalize well to a different region. Here, we leverage the Siamese neural network (SNN) to enhance the generalization of neural network models in discriminating between blasts, collapses, and natural earthquakes under regional shifts. Two distinct data sets are analyzed. The model is trained on a data set from northeastern China and tested on an out-of-region data set from the Inner Mongolia Autonomous Region and Gansu Province. We evaluate the prediction performance of the SNN model against the Convolutional neural network (CNN) model using the K-fold cross-validation technique. Results show that both the CNN and the SNN models achieve highly comparable performance on the in-domain validation data set. However, when applied to the out-of-region test data set, the SNN model with target-region anchors can improve the predicted AUPRC values by 7% and 4% compared with that of the out-of-region CNN model and the CNN model with transfer learning using target-region anchors, respectively. Furthermore, Grad-CAM importance weight analysis shows that the SNN model mainly relies on early-arrival P- and S-wave trains. The study suggests that SNN model with target-region anchors can deliver better generalization and flexibility than the conventional CNN model under regional shifts, which is particularly valuable for regions lacking labeled data sets.