HR: 0800h
AN: NS31A-02 [Abstracts]
TI: Fast Numerically Based Modeling for Ground Penetrating Radar
AU: * Sassen, D S
EM: dsassen@geo.tamu.edu
AF: Texas A&M University, The Department of Geology and Geophysics
Texas A&M University
77843-3115, College Station, TX 7784, United States
AU: Everett, M E
EM: everett@geo.tamu.edu
AF: Texas A&M University, The Department of Geology and Geophysics
Texas A&M University
77843-3115, College Station, TX 7784, United States
AB:
There is a need for computationally fast GPR numerical modeling. This includes circumstances where real time
performance is needed, for example discrimination of landmines or UXO's, and in circumstances that require a
high number of successive forward problems, for example inversion or imaging. Traditional numerical
techniques such as finite difference or finite element are too slow for these applications, but they provide results
from general scenarios such as scattering from very complicated shapes with high contrast. Neural networks
may fit in the niche between analytical techniques and traditional numerical techniques.
Our concept is training a neural network to associate the model inputs of electromagnetic properties of the
background and targets, and the size and shape of the targets, with the output generated by a 3-D finite difference
model. Successive examples from various electromagnetic properties and targets are displayed to the neural
network, until the neural network has adapted itself though optimization.
The trained neural network is now used as the forward model by displaying new input parameters and the neural
network then generates the appropriate output. The results from the neural network are then compared to results
from finite difference models to see how well the neural networks is performing and at what point it breaks down.
Areas of poor fit can be addressed through further training.
The neural network GPR model can be adapted by displaying additional finite difference results to the neural
network, and can also be adapted to a specific field area by actual field data examples. Because of this
adaptation ability the neural network GPR model can be optimized for specific environments and applications.
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 Joint Assembly