Teleseismic Denoiser

Methodology and this package.

Qibin Shi, Marine Denolle. Improved observations of deep earthquake ruptures using machine learning. ESS Open Archive . July 09, 2023. DOI: 10.22541/essoar.168889869.98015524/v1

Datasets

Shi, Q., & Denolle, M. (2023). Training Data of the Waveform Denoiser to separate teleseismic earthquake and noise signals [Data set]. UW photonic sensing facility. https://doi.org/10.6069/69E7-6667

This package uses the auto-encoder neural network to separate teleseismic earthquake and noise waveform recordings. For earthquake studies, this deep neural network can be considered as a multi-scale and multi-task denoiser. We recommend both the direct use of the denoiser and further transfter-learning with your own datasets. The denoiser uses the pretrained kernel WaveDecompNet(Yin et al 2022) to learn high-level features.

Installation

If you like using conda to manage virtual environments,

conda create -n codetest python=3.9
conda activate codetest

To install,

pip install git+https://github.com/qibinshi/TeleseismicDenoiser.git

Configuration file

Download/copy tests/config.ini to your work directory.
It allows you to manage the directories (absolute path!) , data shape, model size and training strategies. Modified the [directories] section to match your computer.
Specify in section [data] and [prediction] whether you would like to run with the demo datasets in the package or the larger dataset (download or prepare your own data).

Run executables to test the package

The default option specified in config.ini is to run the code with the demo datasets and pretrained models in the package.

In your work directory, run the executables for quick tests:

To train the model with demo data

denote_train

To test the pretrained model with demo data

denote_test

To predict from the demo noisy data

denote_predict

The code has been tested on various UNIX machines with both CPU and GPUs. If the code does not run, check your directories (absolute path) and your environment.

Prepare the full dataset for a complete study

You can either download our training data for testing the performance or prepare your own dataset following the data format to do a transfer learning. We use the H5py format for waveform data. The training data is in H5py format with the keys quake and noise. The noisy input data for application is in H5py with a signal key quake.

We process and assemble the waveform data using the FDSN client. See the list of DOIs of the original data. Our datasets can be downloaded with the links provided below.

Training data:

Big dataset of high-SNR teleseismic waveforms for M6+ earthquakes

Key quake: A numpy array with shape of (X, 50000, 3). It represent X recordings of 3-component earthquake waveforms, of which each trace has 50,000 sampling points. Our default sampling rate is 10 Hz.

Key noise: A numpy array with shape of (X, 10000, 3). It represent X recordings of 3-component noise waveforms preceding the P arrival, where each trace has 10,000 points.

Note:

The code split the training data into train, validate and test sets. The executables denote_train and denote_test are both run with the training data.
The names of keys can vary, with config.ini modified accordingly.

Noisy application data

Big dataset of teleseismic waveforms for M6 deep earthquakes including both high- and low-SNR data

Key quake: A numpy array with shape of (X, 3000, 3). It represent X recordings of 3-component earthquake waveforms, of which each trace has 3,000 sampling points. Our default sampling rate is 10 Hz. Hence the length of each waveform is 300 seconds.

xtyangpsp/TeleseismicDenoiser