A Three-component Localization Method for A Single Station Based on Deep Neural Networks

Note

The code is implemented in PyTorch, requiring pre-configuration of the corresponding environment.

Figure 1 localization examples.

This figure displays four predictions from the model on the test set, with the right-side heatmap illustrating the back-azimuth predicted by the classification model.

Figure 2 classification label and classification model

Panel (a) shows the corresponding classification labels, while panel (b) illustrates the architecture and output of the final layer in the classification model under this configuration.

Article abstract

This study proposes a deep neural network-based approach for single-station three-component seismic localization. The method employs two networks that takes three-component waveforms as input to predict epicentral distance and back-azimuth respectively. Completely data-driven, our approach requires no prior conditions or manual intervention. We trained the model using a dataset containing 367k high-quality seismic events, followed by systematic evaluation across three dimensions: (1) performance comparison with mainstream localization models, (2) robustness testing under low SNR conditions, and (3) generalization assessment using K-Net data. Experimental results demonstrate that our method achieves high-precision single-station three-component localization in most scenarios, validating its effectiveness and practical utility.