SurfTomoCNN-new

Hu, J., Qiu, H., Zhang, H., & Ben-Zion, Y. (2020). Using Deep Learning to Derive Shear- Wave Velocity Models from Surface-Wave Dispersion Data. Seismological Research Letters, 91(3), 1738-1751.

Install environment

You should have Anaconda3.4 and Matlab>=2014

pip install requirement.txt

Generate training dataset

cd ./1BuildTrainingTestDataset # see readme and run some scripts to genertate training dataset

Train

cd ./2TrainingAndTestScripts

All training parameters used during training can be modified in config.py

class Config(object):
    def __init__(self):
        self.filepath_disp_training = '../DataSet/TrainingData/0.5km/USA_Tibet/disp_combine_gaussian_map/'
        self.filepath_vs_training   = '../DataSet/TrainingData/0.5km/USA_Tibet/vs_curve/'
        self.filepath_disp_real    = '../DataSet/TestData/real-8s-50s/China/disp_pg_real/' 
        self.batch_size = 64     # training batch size
        self.nEpochs = 600       # maximum number of epochs to train
        self.lr = 0.00001        # learning rate
        self.seed = 123          # random seed 
        self.plot = True         # show validation result during training
        self.alpha=0.0000        # damping, not used here
        self.testsize=0.2
        self.pretrained =True
        self.start=600           # training from 600th model.
        self.pretrain_net = "./model_para/model_epoch_"+str(self.start)+".pth"

For training, please set self.filepath_disp_training and self.filepath_vs_training

python Main_Train.py

Watch training loss and validation loss variation

tensorboard --logdir ./runs

Plot final training and validation loss

After training, then you plot training loss and validation loss for publish

python ./PlotTraingLoss.py # you should change some parameters in this script .

Predicting

For prediction, set your real dispersion curve path self.filepath_disp_real

Set self.start=number, number is your model ID stored in ./model_para/

self.pretrained must be True

The name and content of real dispersion curve file should be like examples

e.g. lat_lon.txt

period1 ph_vel1 ph_un1 gr_vel1 gr_un1
period2 ph_vel2 ph_un2 gr_vel2 gr_un2
....

periodN ph_velN ph_unN gr_velN gr_unN

python Main_Predict.py

Extract Vs map at a depth layer

cd ./ExtractResultMap

see Readme to run some scritps.

copy layers_vs_usa and layers_vs_usa_tibet to ../3GMT_plotResults/

Plot Vs Result

cd ../3GMT_plotResults

See readme to run some scripts to plot results.

Left to right: Test1 (usa data as training dataset), Test2 (usa-tibet data as training dataset), Shen et al. (2016)

By Jing Hu

Date 2020-11-24

Email [email protected]