Array phase detection using TPhaseNet models

Code related to the submitted paper Adapting deep learning phase detectors for seismic array processing.

!!This is work in progress and there may be still bugs an incomplete descriptions. Feel free to reach out.!!

Preparations

Dummy training data from ARCES, SPITS and FINES arrays are provided here on Github. Complete test data, models trained on complete training data, and test data predictions can be found here:

https://doi.org/10.17605/OSF.IO/27FPK

Due to file size limit of 5 GB, some files are split and need to be merged after download:

python split_data_for_repo.py merge --pattern "./Downloads/1statfullarray_2022_single_station_waveforms_000*.hdf5" --output data/1statfullarray_2022_single_station_waveforms.hdf5
python split_data_for_repo.py merge --pattern "./Downloads/array25arces_2022_array_waveforms_000*.hdf5" --output data/arces25_2022_array_waveforms.hdf5

Due to overall size limit of 50 GB, test data predictions for array detection model are not uploaded, but can be produced using test data and model (see below).

Note that dummy data must be overwritten in data/ with downloaded data.

Tested setup for installation of required packages :

conda create -n test python=3.10.12
conda activate test
pip3 install -r requirements.txt

Edit config file to choose type of input data, type of model, etc. Several model config files are also provided.

Note that only config_1stat.yaml includes complete explanations for each parameter.

Single station detection

Train model on single station data (explanation of output files will be added):

python train.py --config config_1stat.yaml

If you want to train on a more complete dataset, download data from link above (same for al other models below).

Create evaluation metrics and plots on test data:

To generate meaningful results, you should download the test data predictions (year 2022) from the data repository and save under output/predictions/.

python evaluate_on_testdata.py --config config_1stat.yaml

Ensemble detection

Apply single station model to all array stations in test data for ensemble detection:

To generate meaningful results, you should download the model trained on multy year data from the data repository and save under output/models/.

python predict_on_testdata.py --config config_1statfullarray.yaml

Combining single station output for ensemble detection is done in evaluation script:

python evaluate_on_testdata.py --config config_1statfullarray.yaml

Beam detection:

Train two beam detection models (vertical ad three-component beams) and evaulate on test data:

python train.py --config config_zbeam.yaml
python train.py --config config_3cbeam.yaml
python evaluate_on_testdata.py --config config_zbeam.yaml
python evaluate_on_testdata.py --config config_3cbeam.yaml

Array detection:

Train the array detection model (ARCES array using Set 2 - see paper) and evaluate on test data:

python train.py --config config_arrayarces_set2.yaml
python evaluate_on_testdata.py --config config_arrayarces_set2.yaml

For meanunful evaluation you need to generate test data predicitons (missing in data repository):

python predict_on_testdata.py --config config_arrayarces_set2.yaml

Detection on contineous data

In the config files only ARCES and 1 hour of processing is chosen. SPITS array can be used for example. Edit and adapt.

Single station detection (on ARA0):

In config_1stat.yaml set:

prediction.predict : True # run on contineous data

prediction.combine_array_stations : False # no post-processing (only done for ensemble and beam detection)

prediction.detect_only : False # If True will not write out continoues detection time series

prediction.stations: ['ARA0']

python predict_contineous.py --config config_1stat.yaml

Ensemble detection:

Do the detection on all array elements with single station detection model:

Im model_configs/config_1stat.yaml set :

prediction.predict : True # run on contineous data

prediction.combine_array_stations : 'stack' # do post-processing (stacking)

prediction.stations: ['ARCES'] # use all ARCES stations

python predict_contineous.py --config config_1statfullarray_cont.yaml

Do not write detection time series for each station, but combine and write only phase detectiobs (only works for stacking option):

prediction.detect_only : True

Beam detection:

Includes combining beams if : prediction.combine_beams : True

python predict_contineous.py --config config_zbeam.yaml
python predict_contineous.py --config config_3cbeam.yaml

Array detection:

Only a model trained on ARCES is available.

python predict_contineous.py --config config_arrayarces_set2.yaml

Evaluation of contineous detection

python evaluate_contineous.py --config <config-file-from-above>