A Fast GPR Numerical Model Based on Machine Learning with Application to Full Waveform Inversion

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

Forward modelling of Ground Penetrating Radar (GPR) is often used to facilitate interpretation of complex GPR data and as a key ingredient of full waveform inversion (FWI) processes. As general 3D full-wave electromagnetic solvers are computationally very demanding routine application of advanced GPR modelling is not popular. A novel concept for creating a fast GPR forward model based on machine learning (ML) concepts is presented. This ML-based model is trained using a dataset obtained from a realistic 3D Finite-Difference Time-Domain (FDTD) gprMax model. The fast model is trained for a specific GPR application that can be easily parametrised and have a somewhat constrained variability. However, the training uses GPR A-Scans obtained from very realistic forward models that include all complex scattering effects and antenna coupling mechanisms. To demonstrate the efficiency of the approach an application, using real GPR data, of the fast forward solver within a FWI process, using a global optimiser requiring a great number of forward model calculations, is presented producing very promising results.
Original languageEnglish
Title of host publication 24th European Meeting of Environmental and Engineering Geophysics
PublisherEAGE
DOIs
Publication statusPublished - 9 Sep 2018
Event24th European Meeting of Environmental and Engineering Geophysics - Porto, Portugal
Duration: 9 Sep 201813 Sep 2018
https://events.eage.org/en/2018/24th-european-meeting-of-environmental-and-engineering-geophysics

Conference

Conference24th European Meeting of Environmental and Engineering Geophysics
CountryPortugal
CityPorto
Period9/09/1813/09/18
Internet address

Keywords

  • GPR
  • Machine learning

Fingerprint

Dive into the research topics of 'A Fast GPR Numerical Model Based on Machine Learning with Application to Full Waveform Inversion'. Together they form a unique fingerprint.

Cite this