nfsi-canada / obstransformer Goto Github PK

OBSTransformer is a transfer-learned seismic phase picker for Ocean Bottom Seismometer data (OBS) adopted from the EqTransformer model. OBSTransformer has been trained on an auto-labelled tectonically inclusive OBS dataset comprising ~36k earthquake and 25k noise samples. OBSTransformer is now integrated with the hands-free earthquake location workflow-LOC-FLOW. You can download the auto-labelled data and noise samples used for model training via this link. Below snippet code demonstrates how to deal with the dataset:

import h5py as hp
import matplotlib.pyplot as plt

dataset = "OBST_training_data.hdf5"
d = hp.File(dataset, 'r')
for item in d['data']:
  if not item.endswith('EV'): continue 
  dd = d.get(f'data/{item}')
  p_arr = dd.attrs['p_arrival_sample']
  s_arr = dd.attrs['s_arrival_sample']
  print(p_arr, s_arr)
  plt.plot(dd)
  plt.show()

OBSTransformer is a variant of EqTransformer optimized for OBS data. Visit the parent repository for detailed installation guidelines.

conda create -n obst python=3.7
conda activate obst
pip install --upgrade pip
pip install obspy
pip install EQTransformer

You may encounter version conflicts between packages, such as numpy or protobuf. Try to install the suggested versions to eliminate the issue.

1. Create a text file including instrument information (see stations.dat)
2. Build the required station metadata using the provided shell script (build_json.sh; needs two addresses)
3. Prepare the dataset directory (see dataset/)
4. Run the detection.py code (python detection.py dataset/)

Niksejel, A. & Zhang, M., 2024. OBSTransformer: A Deep-Learning Seismic Phase Picker for OBS Data Using Automated Labelling and Transfer Learning. Geophysical Journal International, ggae049, https://doi.org/10.1093/gji/ggae049.