Quantitative Inversion of Multiantenna Ground-Penetrating Radar Data with Modeling Error Correction Based on Long Short-Term Memory Cells

Quantitative inversion of GPR data opens the door to precise characterization of underground environments. However, in order to make the inverse scattering problem solution easier from a computational viewpoint, simplifying assumptions are often applied, i.e., two-dimensional approximations or the c...

Full description

Saved in:
Bibliographic Details
Published in:Remote sensing (Basel, Switzerland) Switzerland), 2024-06, Vol.16 (12), p.2050
Main Authors: Fedeli, Alessandro, Schenone, Valentina, Randazzo, Andrea
Format: Article
Language:English
Subjects:
Antennas
Approximation
Dielectric properties
Dimensional measurement
Error correction
Error correction & detection
Finite difference time domain method
Ground penetrating radar
Inverse scattering
Long short-term memory
Memory cells
Neural networks
quantitative imaging
Radar data
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Quantitative inversion of GPR data opens the door to precise characterization of underground environments. However, in order to make the inverse scattering problem solution easier from a computational viewpoint, simplifying assumptions are often applied, i.e., two-dimensional approximations or the consideration of idealized field probes and electromagnetic sources. These assumptions usually produce modeling errors, which can degrade the dielectric reconstruction results considerably. In this article, a processing step based on long short-term memory cells is proposed for the first time to correct the modeling error in a multiantenna GPR setting. In particular, time-domain GPR data are fed into a neural network trained with couples of finite-difference time-domain simulations, where a set of sample targets are simulated in both realistic and idealized configurations. Once trained, the neural network outputs an approximation of multiantenna GPR data as they are collected by an ideal two-dimensional measurement setup. The inversion of the processed data is then accomplished by means of a regularizing Newton-based nonlinear scheme with variable exponent Lebesgue space formulation. A numerical study has been conducted to assess the capabilities of the proposed inversion methodology. The results indicate the possibility of effectively compensating for modeling error in the considered test cases.
ISSN:2072-4292
DOI:10.3390/rs16122050