E3WS: Earthquake Early Warning Starting From 3 s of Records on a Single Station With Machine Learning

Welcome to the E3WS. Do you have a station? Start monitoring earthquakes!

If you use this software, you should cite the code as follow:

**E3WS, Pablo Lara et al. 2023, **

And the paper:

Pablo Lara, Quentin Bletery, Jean-Paul Ampuero, Adolfo Inza, Hernando Tavera. Earthquake Early Warning Starting From 3 s of Records on a Single Station With Machine Learning. Journal of Geophysical Research: Solid Earth.

E3WS article: https://doi.org/10.1029/2023JB026575

Installation

i) git clone https://github.com/PabloELara/E3WS.git

ii) Create E3WS environment

Install miniconda3 (https://docs.conda.io/en/latest/miniconda.html) or miniforge3 for Raspberry Pi 4 (https://github.com/conda-forge/miniforge).
Create the environment: $conda create -n E3WS python=3.7
Activate environment: $conda activate E3WS

iii) Install E3WS dependences ('pip' command is your friend):

python = 3.7 - 3.9
xgboost = 1.6.1
scipy = 1.8.1
python-speech-features = 0.6
scikit-learn = 1.1.1
PyGeodesy = 23.3.23
obspy = 1.3.0

E3WS models

E3WS consists of 3 stages: detection, P-phase picking and source characterization.

For P-phase picking and source characterization (magnitude and location) the models are already defined. You can find them in the models folder. For example MAG7tp3*.joblib, it means the magnitude model uses 7 seconds before P-phase and 3 seconds after.

For detection, you must create your own model with intrinsic noises of the station to be installed. Relax, it is not difficult.

Inside the 'DET/build_DET/' folder:

Have 10 days (or more) of continuous data to extract the noise (we must reach 900000 samples) and add it to the 'data/' folder.
Eliminate the seismic records in these 10 days. I made an automatic program to remove it ('gen_catag.py'). I strongly recommend removing earthquakes from the trace manually, following the format of the '.csv' of the 'picked/' folder.
Generate the feature vector with the program 'pb_FV_noise.py', it will create a csv file in the 'atr_noise/' folder.
Download the earthquake feature vector folder 'atr_eq/' at https://mega.nz/folder/E2UTXIQZ#rH_k9nNrUIgU3D04rzPzzQ
Finally, we have our noise ('atr_noise') and earthquake ('atr_eq') attributes. Now we have to train our detector model! Adapt the file pb_save_DET_model.py, and it will create our model in 'saved_models/'.

Run E3WS

Time to monitor earthquakes. An example of how E3WS works is in the 'real_time/' folder. E3WS detected the M5.6 earthquake of January 7, 2022 in Lima, Peru. The first E3WS estimate was M5.3 based on the first 3 seconds of P-wave. Continuous updates converged to M5.6 and are placed in the results/ folder.