Xiao-Li We, Chun-Xia Zhang, Sang-Woon Kim, Kai-Li Jing, Yong-Jun Wang, Shuang Xu, Zhuang-Zhuang Xie
**This code implemented by Xiao-Li We for 2D seismic fault detection is based on TensorFlow 1.14.0. GPU is RTX 2080 Ti.
##Run an example for fault prediction
If you would like to try predict examples, you can run test_transfer_matrix.py and dowmload the folder[log] to use. Folder data_patches is our test patches.
Two models in folders case1 and case2 are FCFT-Focal loss-Case1 and FCFT-Focal loss-Case2.
**To train our CNN network, with the help of the open source code from Hale ( https://github.com/dhale/ipf), we automatically created 800 pairs of synthetic seismic image and corresponding binary masks. You can find an example in folder data_set, in which f is the seismic image and label is the corresponding binary image.
All images are composed of 200 by 200 pixels. And then we extract patches from them.
The patch training set (train_x_step_3,train_y_step_3; train_x_step_2,train_y_step_2) and validition set (val_x_step_1,val_y_step_1) can be generated by proprecessing.py.
And patch test set (train_x_transfer_3,train_y_transfer_3;test_x_transfer_3,test_y_transfer_3)can be extracted by preprocessing_transfer.py
#training
Run pre_train.py to start the pre_traininging process.
Run transfer_train.py to start the transfer learning stage.
To verify the generalization ability of our model, we also test it with some real data from Netherland offshore F3 block (https://www.opendtect.org/osr/Main/NetherlandsOffshoreF3BlockComplete4GB).
More details can be found in manuscript.