Please refer to this MS thesis for details: https://repository.kaust.edu.sa/bitstream/handle/10754/662251/jamesAudretschThesis.pdf?sequence=5
The idea is to feed in Spectrogram images to a convolutional neural network in order to detect earthquakes.
The final goal is to make a Generalized Earthquake Detector, meaning it doesn't need to have seen data in a location to be able to find earthquakes.
Retrieving and processing the data:
https://github.com/jamesaud/seismic-analysis-toolbox
The data is too large to be held on github, but you can download the data with the toolbox. I will make my dataset available in the future (on Dropbox or something).
On my machine I have roughly 4 million waveforms downloaded!
Written in Pytorch.
If you are familar with Pytorch, the main code will look familiar:
https://github.com/jamesaud/earthquake-detector-9000/blob/master/main.py https://github.com/jamesaud/earthquake-detector-9000/blob/master/pytorch_utils/utils.py
The configuration is in json (or dict in Python):
https://github.com/jamesaud/earthquake-detector-9000/blob/master/config.py
In the 'main' code, if the configuration is set to 'environment', configuration = 'environment' , then it will read in the configuration from validator/config.json.
python main.py runs the neural network.
Use this model for best performance and efficiency:
https://github.com/jamesaud/earthquake-detector-9000/blob/master/models/mnist_3_component.py
It's an interesting model because it feeds 3 components (N, Z, E), meaning three spectrograms, and then combines features to make a detection.
The convnet does 99.0 - 99.9% accuracy on single location QUALITY datasets.
Many times, the automatic generation of seismic datasets has improperly labeled events or noise. In order to determine what is good quality, use Data Validation.
To run the conv net on every location, run:
python data_validate.py
to produce results that tell you the detection accuracy for each location:
I kept locations that were 97% accurate and higher from the Data Validation step. I used this as my final dataset to train a generalized earthquake detector.
Run docker-compose up to get Tensorboard running on localhost:6000
The code in main is configured to write to the folder for visualization.
There are some tests written:
https://github.com/jamesaud/earthquake-detector-9000/tree/master/tests
This works with cuda 7.5, because that's what my workstation ran, and required PyTorch version 0.3.x. Now, if you upgrade to PyTorch 0.4 or 1.x there are some breaking changes. It took around 10 minutes to refactor the code after trying to run on version 0.4.x to work again.
Have anaconda installed and run the following command:
conda env create -f environment.yml
enter the new environment
conda activate earthquake
now you are set to run the code:
python main.py
(This won't work until you specify where spectrograms are in the config.py file)
This is a work in progress still. Some of the code does bad practice/unconvential things.