HR: 0800h
AN: NS31B-0392    [Abstracts]
TI: 3-D ADI-FDTD modeling of GPR backscatter from complex targets for the training of artificial neural networks
AU: * Sassen, D S
EM: dsassen@geo.tamu.edu
AF: Texas A&M University Dept. Geology and Geophysics, Texas A&M University Dept. Geology and Geophysics, College Station, TX 77843 3115, United States
AU: Everett, M E
EM: everett@geo.tamu.edu
AF: Texas A&M University Dept. Geology and Geophysics, Texas A&M University Dept. Geology and Geophysics, College Station, TX 77843 3115, United States
AB: Artificial neural networks can provide approximate solutions to ground-penetrating radar (GPR) problems in cases where real time performance is needed. Examples include discrimination of landmines or UXO's, and in circumstances that require a high number of successive forward problems, for example inversion or imaging. The training of neural networks to work within even a limited range of targets and electromagnetic properties requires a large set of successive examples generated from numerical methods such as finite difference time domain (FDTD). The traditional FDTD technique suffers from numerical dispersion unless time steps are kept below the Courant stability limit. The accurate modeling of electromagnetic scattering by complex targets require a refined grid, subgrids, or conformal grids that can significantly increase computation time, making neural network training inefficient. A relatively recent FDTD technique, ADI-FDTD, uses implicit equations that help to cancel numerical dispersion and allow for unconditionally stable modeling of EM propagation and therefore is not bound by the Courant stability limit. The technique is especially efficient for the accurate modeling of complex targets. Our ADI-FDTD code includes the ability to refine the model grid and to implement a conformal gridding to improve model accuracy without effecting the overall computation time. We will explore the tradeoff in computation time and accuracy in modeling the GPR backscatter of various targets using both the ADI-FDTD technique and the traditional FDTD technique for the purpose of neural network training.
DE: 0545 Modeling (4255)
DE: 0555 Neural networks, fuzzy logic, machine learning
DE: 0669 Scattering and diffraction
DE: 0689 Wave propagation (2487, 3285, 4275, 4455, 6934)
SC: Near-Surface Geophysics [NS]
MN: 2007 Fall Meeting