Reference: Hu et al., (2019) Using deep learning to derive shear wave velocity models from surface wave dispersion [submitted to SRL]
This repository is used to store scripts and an example of surface tomography given real data set
The dataset and scripts are the part of the paper (section 3.1) for the continental China case.
- You should install Pytorch 0.4 version and Anaconda3.4 and tensorboardX and basemap (for plotting Vs).
conda install pytorch=0.4.0 cuda90 -c pytorch; pip install tensorboardX;
For basemap installation see https://matplotlib.org/basemap/users/installing.html
All python scripts are py3.
cat Dataset.tar.gz.* | tar -zxv
Training dataset: USA type (~7000 1-D Vs models from the USA (Shen et al., 2013)); USA-Tibet type (USA type + ~640 Tibet models)
Test dataset: ~4000 pairs of disperion images associated with phase and group velocity (8-50s). (Shen et al., 2016)
-
scriptsUSA: using only USA type as a training dataset to train and then using test dataset to predict 1-D Vs models -
scriptsUSATibet: using USA type plus ~640 Tibet models as a training dataset and then using test dataset to predict 1-D Vs models
For both tests, if you train again, run Main_train.sh. If you test, run Main_test.sh
The trained model at 600th epoch is included in model_para
PlotVsis used to make a comparison between cnn-based Vs models and results of Shen et al. (2016)
It would read ./scriptsUSA/vs_cnn and ./scriptsUSATibet/vs_cnn as well as ./PlotVs/data/vs_sws_China to plot