/SFD-CNN-TL

Seismic fault detection in real data using Transfer Learning from a Convolutional Neural Network pre-trained with synthetic seismic data -- Authors: Augusto Cunha, Axelle Pochet, Helio Lopes, Marcelo Gattass

Primary LanguageJupyter NotebookOtherNOASSERTION

This page is available as an executable or viewable Jupyter Notebook: Binder NbViwer

This code allows to reproduce results of our article:

Seismic fault detection in real data using Transfer Learning from a Convolutional Neural Network pre-trained with synthetic seismic data https://doi.org/10.1016/j.cageo.2019.104344

Publications

If you find this work helpful in your research, please cite:

@article{cunha2020SFD,
title = "Seismic fault detection in real data using transfer learning from a convolutional neural network pre-trained with synthetic seismic data",
journal = "Computers & Geosciences",
volume = "135",
pages = "104344",
year = "2020",
issn = "0098-3004",
doi = "https://doi.org/10.1016/j.cageo.2019.104344",
url = "http://www.sciencedirect.com/science/article/pii/S0098300418307040",
author = "Augusto Cunha and Axelle Pochet and Hélio Lopes and Marcelo Gattass",
keywords = "Transfer learning, Convolutional neural network, Seismic fault"
}

Notebook visualization

We provide two option to easily open and see our ipython notebooks:

We strongly recommend using NbViwer to visualize our notebooks instead open in GitHub, because you will see the interactive plots.

Dependencies

environment.yml

To install all requirements in the environment use:

conda env create -f environment.yml

Sometimes is needed to create a kernel of the environment to jupyter notebooks:

python -m ipykernel install --user --name sfd --display-name "SFD-CNN-TL"

Our F3 Block pre-trained model can be found in:

base_model/model.json
base_model/model.h5

Our F3 Block slice manualy interpreted dataset can be found in:

dataset/fault
dataset/nonfault

Transfer learning methods:

ft.py : full fine tuning (FFT)
mlp.py : feature extractor with Multi Layer Perceptron (FE-MLP)
svm.py : feature extractor with Support Vector Machine (FE-SVM)

Default parameters are set to produce the results presented in the article.

Generated models can be saved by setting the boolean value save=true in functions create_model(). they will be save in the output/ directory.

Classification results:

classify.py : generates a classification file for models saved as .json and .h5
classify_with_SVM.py : generates a classification file for models saved as .pkl

Classification files is saved in directiry classification/output/ It contains patches coordinates associated to a class value (1 for fault, 0 otherwise)

We provide a region of a real section where a fault is clearly visible as demo in the classification/ directory. Other sections can be classified modifying the classify.py and classify_with_SVM.py files.

Metrics:

metrics.py : computes quality metrics (accuracy, sensitivity, specificity = recall, F1-score, ROC AUC and we added precision)

Interactive examples:

classifyAndViewF3.ipynb: Classify selected sections of F3 seismic data with all pretrained methods and show the results
classifyAndViewGSB.ipynb: Classify selected sections of GSB seismic data with all pretrained methods and show the results
TrainAndSave.ipynb: Train all methods with one interpreted section of a real data and save network weights