GPRNet - GPR Inversion Using Deep Learning

Update 1/1/2022: This repo is now updated to reflect the figures and results that are produced in the paper. If you downloaded the files before 1/1/2022, please re-download again.

Note: This repository contains only codes. To reproduce the figures at a local machine, please download all data sets here: https://bit.ly/36hDrRj (The password is a u t o m a t e G P R , without spaces)

Key Points from paper:

• We propose a deep learning-based EM velocity inversion for GPR zero-offset data
• Tests on synthetic examples show accurate velocity inversion results
• Applications to field data yield predictions that agree with the velocity models derived from previous physics-based inversion studies

GPRNet Architecture comment:

GPRNet architecture is found at DLcodes/GPRNet.py

• Essentially, it's an encoder-decoder based Convolutional Neural Network (CNN).
• The framework is designed based on the on DeepLabV3 architecture(https://arxiv.org/abs/1706.05587)

Synthetic EM Velocity Inversion

1D Scenario:-

1. Run 1-Generate_1D_models.py

   • Generates 10,000 random velocity profiles
   • Creates these files:
     • Synthetic/Data/1D/ep.mat (dielectric permittivity for FDTD simulation to obtain GPR data)
     • Synthetic/Data/1D/veltd.npy (raw velocity files)

2. Run 2-FD_GPR_sim.m

   • Perform FDTD on ep.mat to create raw 1D GPR gathers
   • Creates Synthetic/Data/1D/fdrawgather.mat

3. Run 3-Process_1D_GPRtraces.py

   • Removes first arrivals and remove any data pair that has NaN values
   • Creates these files:
     • Synthetic/Data/1D/xTrain_gathers.npy
     • Synthetic/Data/1D/yTrain_vels.npy

4. Run 4-Data_Loader.py

   • Split data into training, testing, and validation
   • Creates these files:
     • Synthetic/Data/1D/ForDL/Synthetic_Xtrain_1d.npy
     • Synthetic/Data/1D/ForDL/Synthetic_ytrain_1d.npy
     • Synthetic/Data/1D/ForDL/Synthetic_yvalid_1d.npy
     • Synthetic/Data/1D/ForDL/Synthetic_xvalid_1d.npy
     • Synthetic/Data/1D/ForDL/Synthetic_Xtest_1d.npy
     • Synthetic/Data/1D/ForDL/Synthetic_yTrue_1d.npy

5. Run 5-Synthetic_1D_DL_Training.py

   • Trains GPR-Velocity
   • Creates these files:
     • Synthetic/Weights/weight_GPRNet_n16k20.h5
     • Synthetic/Weights/weight_GPRNet_n16k20.csv

6. Run 6-Synthetic_1D_DL_Prediction.py

   • Applies trained weights to testing data set
   • Creates these files:
     • Synthetic/Data/1D/ForDL/Synthetic_ypred_1D.npy

7. Run 7-Synthetic_1D.ipynb (jupyter notebook)

   • Reproduces Figure 3 and Figure 4

2D Scenario :-

8. Run 8-Process_2D_models.py

   • Synthetic/Data/2D/yTrue2D_vel_dd.npy is the given 2D velocity model (in m)
   • This script converts the velocity model into time depth domain and to dielectric permittivity to be used for 2D common-offset GPR FDTD simulation
   • Creates these files:
     • Synthetic/Data/2D/yTrue2D_ep.mat; (for FDTD simulation)
     • Synthetic/Data/2D/yTrue2D_vel_td.npy; (for ground truthing predicted velocity model)

9. Run 9-FD_sim_2Dtestingmodel.m

   • Simulates GPR data (common-offset)
   • Creates Synthetic/Data/2D/fdraw_2D.mat

10. Run 10-Synthetic_2D.ipynb (jupyter notebook) - Part 1

    • Processes and predicts velocity model from fdraw_2D.mat
    • Creates these files:
      • Synthetic/Data/2D/ypred2D.npy
      • Synthetic/Data/2D/ep_ypred2D.mat (used for forward data to see data matching)

11. Run 11-FD_sim_testingmodel_ypred.m

    • Creates forward data from prediction (ypred2D.npy)
    • Creates Synthetic/Data/2D/fdraw_predicted_data.mat

12. Run 12-Synthetic_2D.ipynb (jupyter notebook) - Part2

    • Reproduces Figure 5 and 6

Field Application

13. Run 13-generate_vel.py

    • Creates these files:
      • Field/Data/ep.mat
      • Field/Data/veltd_raw.mat

14. Run 14-GPR_sim.m

    • Simulates 50,000 GPR traces (this is a large job, might want to split this into a few parts)
    • Creates these files:
      • Field/Data/fdrawgathers.mat (intermediate files, skipping upload)

15. Run 15-process_GPR_Vel_part1.m

    • Preprocess GPR data and Velocity
    • Creates these files: (intermediate files, skipping upload)
      • Field/Data/AllRawGathers.mat
      • Field/Data/veltd_raw_corr.mat

16. Run 16-process_GPR_Vel_part2.py

    • Create data and velocity for GPRNet training
    • Augmentation of data set takes place here
    • Creates these files: (intermediate files, skipping upload)
      • Field/Data/ForDL/GPRData.npy
      • Field/Data/ForDL/Vel.npy

17. Run 17-field_Data_Loader.py

    • Splits data set into training, testing and validation
    • Creates these files:
      • Field/Data/ForDL/field_X_train.npy
      • Field/Data/ForDL/field_X_valid.npy
      • Field/Data/ForDL/field_X_test.npy
      • Field/Data/ForDL/field_y_train.npy
      • Field/Data/ForDL/field_y_valid.npy
      • Field/Data/ForDL/field_y_true.npy

18. Run 18-Field_DL_Training.py

    • Trains GPR data and Velocity
    • Creates these files:
      • Field/Weights/weight_GPRNet_n32k10.h5
      • Field/Weights/weight_GPRNet_n32k10.csv

19. Run 19-extract_codata.m

    • Reads wurtsmith_line1.sgy field data and extracts common-offset data
    • Creates Field/Data/rawfielddata/codata.mat

20. Run 20-Field_Application.ipynb (jupyter notebook)

    • Reproduces Figure 7, 8, 9, 10
    • Applies trained weights to common-offset data to obtain field prediction
    • Follow instructions inside notebook to simulate forward data based on field prediction
    • Creates these files:
      • Field/Data/pcsfielddata/ProcessedFieldData_rev.npy
      • Field/Data/prediction/FieldPrediction_rev.npy
      • Field/Data/prediction/ep_FieldPrediction_rev.mat
      • Field/Data/prediction/rawgather_Stacked1D_fieldprediction_rev.mat (this is created in 20a-FD_pred.m)
      • Field/Data/prediction/ForwDataFrPred_rev.npy