ArrayNet: A combined seismic phase classification and back-azimuth regression neural network for array processing pipelines, 2023.

Code and examples related to the paper ArrayNet: A combined seismic phase classification and back-azimuth regression neural network for array processing pipelines.

To train an ArrayNet model (super-model) for the ARCES array run :

bash run.sh

Edit the script and adjust to your working environment

Input data for super-model in tf/data:

• times_merged_arces_4Fre.npy : time stamp for seismic arrivals
• X_merged_arces_4Fre.npy : co-array phase patterns for all arrivals
• y_cl_merged_arces_4Fre.npy : Arrival label (phase type)
• y_reg_merged_arces_4Fre.npy : Back-azimuth to event source

Input data for super- and sub-model in tf/data:

• times_merged_arces_4Fre_regional.npy : time stamp for seismic arrivals
• X_merged_arces_4Fre_regional.npy : co-array phase patterns for all arrivals
• y_cl_merged_arces_4Fre_regional.npy : Arrival label (phase type)
• y_reg_merged_arces_4Fre_regional.npy : Back-azimuth to event source

Edit `experiments.py' to train one of both models.

Due to limitations on file size on github we provide a reduced data set for training. However, we also provide the model trained with the full data set in (super-model only for now):

tf/output_full/

The models trained with the reduced data set which can be reproduced are in:

tf/output/

Call this script to evaluate the model trained with the full data set, including confusion matrix, classification metrics, and back-azimuth residuals (super-model only for now):

python evaluate_models.py

Call this script to generate new input data from ARCES array data (saved under tf/data/ with basename test ):

python generate_input_data.py

Related publication

• Andreas Köhler, Erik B. Myklebust. ArrayNet: A combined seismic phase classification and back-azimuth regression neural network for array processing pipelines. Accepted for publication in BSSA, 2023, https://doi.org/10.1785/0120230056

License

See LICENSE.txt for more information.

(back to top)

Contact

Andreas Köhler - andreas.kohler@norsar.no - ORCID

Erik B. Myklebust - ORCID

Project Link: https://github.com/NorwegianSeismicArray/arraynet

(back to top)

Acknowledgments

• ArrayNet models are built with TensorFLow
• ARCES waveform data from which the input data was generated are available via the Norwegian EIDA node
• Reviewed seismic event bulletins from which the input data labels were obtained are available from the Finish National Seismic Network and NORSAR

(back to top)