Computer Simulation of Ground Penetrating Radar Propagation Through A Freshwater

You might also like

Download as doc, pdf, or txt
Download as doc, pdf, or txt
You are on page 1of 22

 

 
 
 
 
Computer Simulation Of Ground
Penetrating Radar Propagation
Through A Freshwater/Saltwater
Interface
 
James A. Allen
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Abstract

A two-dimensional finite-difference time-domain model was used to simulate

the propagation of radar waves through a freshwater/saltwater interface. We assume

the interface has a conductivity that increases linearly with depth. Ground conditions

roughly similar to those in Key Largo, Florida were simulated. Numerical instabilities

presented a major problem in simulating radar propagation under these conditions. A

majority of the work comprised of attempting to remove the numerical instabilities, or

“ringing,” from the simulation. The amplitude of this “ringing” in many models was

as large as the simulated waves derived from the freshwater/saltwater interface.

Several methods were attempted to reduce the ringing. Altering the algorithm for the

source of the wave and using post-run spatial filtering came the closest to eradicating

the ringing, but the ringing still appeared after ~60 ns of simulated wave travel time.

Variations in the gradient of the interface layer impact the radar wave

propagation in several ways. As the gradient increases, the amplitude of the radar

wave reflected off the interface increases. At large gradients, the rate of increase in

amplitude becomes less. The half width of the negative pulse decreases exponentially

with increasing gradient. Finally, the angle at which the incoming waves strike the

interface layer does not significantly affect the shape of the waves.

Introduction

This thesis will assess the conditions under which a reflection from a

freshwater-saltwater interface and associated layers may be discernible in ground

penetrating radar records. A finite-difference time-domain (FDTD) model is used to

simulate the conditions analogous to those at Key Largo, FL, the northernmost island
of the Florida Keys. The water table on the island is close to sea level, and thus varies

in depth with land elevation. In the northern-central sections of the island, the land

elevation ranges from 3 to 4 meters, and 1 to 2 meters of freshwater overlies brackish

to saline water. However, the elevations near the coast are less than 2 meters, and

there the freshwater layer pinches out. Thus, other than the interior of the island, the

freshwater lens is small to non-existent.

Both GPR profiles and core samples were taken from a site on north-central

portion of Key Largo (Figure 1). There, the water table lies at 3 meters depth, with 1-

1.5 meters of freshwater overlying a freshwater/saltwater interface zone 2 meters

thick (Inman, 1999). At these sites, ground conductivities are 1-2 mS/m above the

water table, increasing to close to 100 mS/m at the base of the interface (Inman,

1999). GPR surveys show several distinct arrivals. Arrivals at 70 ns and 140 ns were

interpreted as the water table and a multiple of this water table. An arrival at 90 ns,

corresponding to a depth of 4 meters, may correspond to the top of the

freshwater/saltwater interface. A strong conductivity gradient in this interface may be

responsible for the arrival seen in the GPR studies. This arrival is seen throughout the

GPR record (Figure 2).

Ground Penetrating Radar

Ground penetrating radar (GPR) has been used extensively in the geologic

community for quite some time, but usage escalated during the 1980s. With studies in

geology, GPR antennae frequencies that are less than ~500 MHz are typically used

(Reynolds, 1997).
Figure 1 – The study area in Key Largo, Florida. GPR data came from sites marked A-D.

Figure 2 – Typical GPR response collected from Key Largo, Florida. The interface reflection can been
seen as the arrival at ~90 ns.
 
Most of the first work in GPR studies was related to polar ice sheet experimentation.

In the present day, GPR is used in many different geologic, engineering, and

archaeological applications.

A typical GPR system is comprised of several parts: a signal generator, which

produces the wave that will propagate through the medium, transmitting and receiving

antennae, and a recorder, to store the signal recorded at the receiver. The wave’s

travel time is very short, usually on the order of hundreds of nanoseconds. Obviously,

very sensitive instrumentation is required to differentiate between signals received in

relatively short time intervals. The antennas themselves can be set in two modes:

monostatic and bistatic (Reynolds, 1997). In monostatic mode, one antenna acts as

both transmitter and receiver, whereas bistatic uses separate antennas. Some systems,

such as the PulseEKKO system used in the Key Largo experiments, specifically use

only bistatic modes.

The transmitter sends many wave pulses into the ground, and the receiving

antenna scans at a fixed interval, around 30 scans per second (Reynolds, 1997). As the

receiver is moved along the ground surface, the recorder stores all the data received,

and the resulting “radargram” is stored digitally. For imaging targets of interest, a

suitable pulse length must be used. This value chosen for the wave pulse is a function

of the hypothesized ground conditions and properties of the GPR imaging equipment

being used. In conducting GPR studies, it is important to choose a correct frequency.

Low frequencies are used to explore greater depths, but are poorer in resolution.

Higher frequency waves do not penetrate the ground as far, but do provide increased

spatial resolution.

The propagation of the radar waves is determined by Maxwell’s equations.

Because of this fact, the behavior of the radar waves can be theorized and predicted.
The speed at which electromagnetic waves propagate through a medium is dependent

on the speed of light, the relative magnetic permeability, and the relative dielectric

constant. These parameters are related to the lithology of the rock matrix and to the

water content of the rock. A media such as air can lead to extremely fast propagation,

and a layer with a large conductivity, for example, can totally absorb the wave energy.

The boundaries between lithologically different layers can be imaged by receiving

reflections off these contacts. The variation between the relative dielectric constants

of adjoining layers dictates how much of the wave energy is reflected, and thus how

prominently the reflected wave arrival appears on the GPR recorder and the resulting

plots.

The signal strength decreases during propagation, as a function of geometric

spreading. In addition, at every interface, energy is partitioned as the wave reflects

and refracts at the contacts. If objects exist in the survey area that have dimensions

that are similar to that of the radar waves, the phenomenon called Mie scattering

occurs, which randomly scatters the energy, and introduces noise in the radargram

(Reynolds, 1997). Finally, energy is attenuated by absorption in conductive media.

Attenuation is dependent on the electric, magnetic, and dielectric properties of

the medium, as well as the frequency of the wave. Attenuation is often described in

terms of skin depth, where skin depth is equal to the depth at which the signal has

decreased in amplitude by 37%, which is equal to 1/e. However, skin depth cannot be

used to determine the depth of penetration of the radar wave. The total path loss,

which is equal to the energy lost over any given distance, can be determined by

computing antenna losses, transmission losses between the air and the ground, losses

caused by the geometrical spreading of the radar beam, attenuation within the ground

as a function of the material properties, and losses due to scattering of the radar signal
from the target itself (Reynolds, 1997). If the reflector is a smooth plane, the incident

signal is similar in nature to the source, but reduced in power. It has been shown that

attenuation is directly proportional to frequency.

The dielectric property of a material is determined by its complex permittivity

and complex conductivity. The relative dielectric constant varies from 1 in air to 81 in

water, with most geologic materials in the range of 3 to 30. However, since most

geologic bodies are a mixture of several minerals and other constituents such as water,

the dielectric constant for a rock in a surveyed area can differ from a published value

of that material. Porosity greatly affects the properties of the terrain, since the

dielectric properties of air and water differ considerably from those of rock material.

The examination of the porosity and permeability of each component of a rock can

lead to an overall determination of the dielectric properties.

The FDTD Model

The two-dimensional FDTD model provides a computer simulation to produce

GPR data, which can then be compared to field data. The propagation of radar waves

in the model is controlled by the layer characteristics, grid resolution and size, and

initial values for the wave, such as the spread time and spatial dimension, all of which

can be manipulated by the user in the computer simulation. These values can be

changed to slightly alter the subsurface, and produce situations similar to field data or

to conduct theoretical experiments. Any combination of altering these values can

drastically change the properties of the waves, including the wavelength, frequency,

velocity, and attenuation. The models values were set to correspond closely to those

conditions found in Key Largo, and then manipulated to investigate how variability in
these parameters would affect the resulting simulated radar response. Understanding

this relationship could help in interpreting the actual GPR data found in Key Largo.

The model computes the values of the out-of-plane component of the electric

field along with the in-plane component of the magnetic field at all points on a two-

dimensional grid for successive time intervals. Iterations and total time elapsed are

specified by setting the time interval (dt) of each iteration, and the total number of

iterations for a run. For example, to produce a run that encompasses 40 ns and using a

dt of 0.25, one needs to run the model for 160 iterations.

The electric and magnetic field values are computed at nodes on a two-

dimensional grid. A finite differencing scheme is used to solve the equations for

electromagnetic wave propagation, implementing Maxwell’s equations. In the model

shown here, a grid of 800 by 800 nodes was used. This nodal area leads to rather

lengthy computations, on the order of 30 to 45 minutes, since the algorithm must

compute the correct amplitude values for 640,000 nodes each iteration.

The vertical and horizontal spacing between nodes (dx and dy) are set by the

user. For a square grid that measures 5 meters by 5 meters, one can use a 200 by 200

nodal grid with dx=dy=0.025 m, or a 100 by 100 nodal grid with dx=dy=0.05 m.

Obviously, since the 100 by 100 grid has 4 times fewer nodes, the calculations for this

situation would be much faster. Unfortunately, there is a practical upper limit as to

how large dx and dy can be. If one chooses too large of a dx or dy value, the

algorithm may not calculate the wave propagation with sufficient accuracy, and the

simulation will fail. In some cases, an instability forms, and the simulation produces

an expanding area of infinitely negative wave energy, which eventually engulfs the

entire experimental grid area. This phenomenon destroys any attempt at gathering

appropriate data. The choice of dx and dy is related to the value chosen for dt. If the dt
value chosen for the model is too large, the algorithm cannot correctly determine the

locations of all waves in a specific time in the future. In practice, when a simulation

becomes unstable, it can be difficult to determine which value (dt or dx) is causing the

downfall of the model. Thus, experimentation to ascertain reasonable values for dx,

dy, and dt is essential.

To simulate a transmitted GPR wave with a two-dimensional model, the

source can be a line source or plane wave. The plane waves are simpler and are useful

for comparing the model with analytical solutions. However, the line source was

chosen to more closely approximate a GPR transmitter, such as was used in Key

Largo. The initial location of the transmitting antenna, or wave source, can be set,

simulating real world conditions. In the Key Largo model, the starting point for the

wave was placed just below the surface (if the wave were allowed to propagate from

the ground surface, the air layer would incur an instability, similar to the situation

found if a large dx or dy is chosen. Since waves travel through the air layer at much

faster velocity than the ground layers, the failure will initiate where the waves start to

separate). Individual properties related to wave propagation can be set as well,

including spread in time of the incident wave (sigt), the spatial dimensions of the

source in node units (srcd), and the amplitude of the source


Figure 3 – Two dimensional plot after 72 ns in the FDTD program. The model properties can be found
in Table 1.
(vvmx). A smaller value for sigt produces a wave with smaller wavelength and srcd

controls the “tail” of the wave, affecting the symmetry of the wave.

Wave propagation is graphed in real time 2-D and 1-D plots (Figures 3 and 4).

The model allows for the placement of GPR “receivers” in any location within the

grid area. The simulated arrivals at “receiver” points can be extracted for comparison

with GPR plots (Figure 4). The propagation of waves through a model with an

interface and freshwater saturated layer can be seen in Figure 5. These real-time plots

indicate to the user whether computations are progressing as intended.

Figure 4 – One dimensional plot displayed during FDTD simulation.


Figure 5 – The FDTD two-dimensional plot through time. This model is comprised of a freshwater
saturated layer, above a 3 meter thick interface layer. The receivers and source lie 2 meters above the
interface layer. The location of the receivers can be seen in Figure 9.
 
Problems with the model

The FDTD model has significant numerical instabilities. A distinct “ringing”

was

found during the two-dimensional modeling (Figure 6), which drastically muddled the

Figure 6 – Ringing found during runs of FDTD simulation. Layers are the same as in Figure 3.
data “received” at simulated receiver locations, and made analyzing the data for low-

amplitude returns almost impossible. To accurately model the conditions at Key

Largo, and to derive correct conclusions about wave propagation through the

interface, the ringing has to be reduced. This experience encompassed four months of

the study.

The first proposed solution was the altering the nature of the wave. The

constants in question here are sigt, the spread in time of the direct wave, srcd, the

spatial dimensions of the source in grid units, and vvmx, the amplitude of the source.

Although changing these values did change the shape of the direct wave, no changes

to these constants brought about a cleaner situation. We supposed a change in the grid

size or the time step might eliminate some noise in the function. The actual

dimensions of the grid were changed, but if a larger or smaller grid was used, the

noise was still present. In conjunction with this, the time step with which the

simulation ran was also changed. It was found that there is an upper limit above which

a larger difference in time between iterations will cause the simulation to fail, but

even an extremely small time step would produce noise.

It was thought that a possible source of the noise was that the individual nodes

comprising the grid were spaced too apart, and that out of this large distance between

these data points arose the noise. A more “concentrated” grid would pack the nodes

closer together, and hopefully allow the wave to pass cleanly from adjacent nodes.

Unfortunately, even with large grid resolutions such as fifty to one hundred nodes per

meter, the noise was still resident. To assist with computational speed on the large

grids, the saltwater zone layer beneath the interface layer was removed, and

consequently, the grid area was reduced. This removal was possible because the

portion of the wave that was refracted at the interface boundary was totally absorbed
over the distance of the layer, and an additional layer further in the ground was not

needed. This reduced computational time, but not the noise.

A close analysis of the noise within the program revealed that the noise never

exceeded the wavelength of any reflected pulse that the receivers could eventually

record. Because of this new revelation, a spatial filtering model was proposed. The

code was altered so that, during the calculations of the FDTD model, the amplitude

value of every node in the grid space was recomputed as the average of the ten

surrounding nodes and itself. This led to very lengthy computations, and effectively

doubled the time over which one FDTD model was completed. The primary difficulty

in choosing an interval for the spatial filtering was the preservation of the reflected

pulses: the spatial filtering must not eliminate this important data. The result of the

spatial filtering was two-fold: a cleaner function was found, but slight alteration of the

reflected pulses was also present. Thus, because the wave properties that possibly

were altered could not be determined, spatial filtering during experimentation was

abandoned.

A primary source of error arises from simulating a circular pulse on a square

grid. The algorithm possibly cannot model this situation perfectly accurate. We

developed an improved circular source, in which the program would “force”

propagation to a certain time, after which the finite difference algorithm would take

over. While the noise did not appear while the algorithm was not controlling the

model, after the time interval was passed, the noise again appeared. Additionally,

while the program was forcing the wave, the calculations took many computer cycles

to complete, and this caused the time to complete the simulation to increase

dramatically. Since the noise still resulted, the


Figure 7 – GPR plot before post-run spatial filtering. The two-dimensional plot is shown at 84 ns.

reformatted circular pulse was not practical.

All of the proposed solutions up to this point had failed, so the next attempted

course of action was taken post-run. The spatial filtering was applied after an entire

run had been simulated. A run before spatial filtering can be seen in Figure 7. The

data were spatially filtered in an outside program, which would complete the task

similar to that done previously within the program. The amplitudes for each receiver

were recomputed as the average of the amplitudes for ten adjacent time intervals. The

result of this filtering can be seen in Figure 8. The ringing can be seen, but the “real”

pulses can be singled out. This method yielded the “cleanest” signal to date.

Finally, new code was written with an improved algorithm. The pulse was

initially propagated some initial distance and established over a circular region before

finite difference calculations were begun. The grid resolution could be decreased
dramatically: consequently, the grid size could be reduced, and the calculations

proceeded much more quickly. This new program was not perfect, however. After a

Figure 8 – GPR plot after post-run spatial filtering: compare to Figure 7. The ringing can still be seen
in the data.
certain time, ringing that was different than that found in the first runs was discovered

(Figure 9). Fortunately, since this ringing seemed to appear after ~70 ns, if the

important arrivals were simulated before this time, the ringing could be ignored. To

produce this situation, the interface was placed close to the receivers and to the

starting point for the wave. Thus, the interface reflection partially overlapped the

direct wave at the “receiver” locations. To extract the interface reflection, the direct

wave was subtracted.


Figure 9 – Ringing seen during run of new FDTD algorithm. Picture is taken while the simulation is
running. Note that the ringing is not as pronounced as in Figure 6, but it definitely present.
Layer Depth Below Surface Dielectric Constant Conductivity (mS/m)
Air 0.0 m 1.0 0.0
Unsaturated 0.0 – 2.9 m 10.0 1.000
Freshwater Sat. 2.9 – 4.2 m 14.0 2.000
Interface 4.2 – 7.2 m 14.0 2.000 (top) – 300.000
Table 1 - Properties of the layers used in the FDTD model. In all cases, the magnetic permeability is
1.0.
In a final attempt to reduce the noise, the reflections at several closely spaced

receivers were “stacked,” or summed after correcting for the difference in arrival

times. In theory, a very large interface reflection should result, and the noise later in

the record would be eliminated, since, from the two-dimensional plots, it appeared

that the noise was moving across the receivers. However, after this idea was

implemented, it was found that the noise was not eliminated, and the eventual

decision was to accept the data as received from the closely placed interface layer. Of

all the trials conducted to attempt to find a clean signal, this method gave the best

results.

The Interface

The layer that holds the most interesting properties is the interface. Since this

layer has a conductivity that varies over depth, several questions arise as to the nature

of the GPR response to this layer. How does the conductivity gradient affect the

amplitude and period of the reflected wave? In addition, does the incident angle of the

direct wave have an effect on the form of the reflected pulse? If the conductivity

gradient has a strong impact on the GPR response, could the gradient be estimated

from real GPR data collected in the field?

The models used to simulate the Key Largo setting consist of 4 layers, with

electrical and magnetic properties as shown in Table 1. The top most layer replicates
air. A 2.9 meter thick layer of unsaturated material lies at the ground surface, below

which is

Figure 10 – The direct wave. The two dimensional model at 44 ns is shown. The model is
composed of just one layer: the freshwater saturated zone: the properties are given in Table 1.
 
the water table, the contact between this layer and the 1.3 meter thick freshwater

saturated layer. Below the freshwater saturated layer lies a layer with a conductivity

that increases with depth, referred to as the interface layer. Three specific parameters

were set for each layer: magnetic permeability (mm), the dielectric constant (ee), and

conductivity (ss), measured in mS/m, which can be varied linearly with depth.

The interface layer was proposed to have a conductivity of 2 mS/m at the top

of the layer, and 300 mS/m conductivity 3 meters below the bottom of the layer. This

situation would produce a gradation of 99.33 mS/m per meter. Several variations on

this scenario were proposed, ranging from 9.33 mS/m2 to 9999.33 mS/m2. The FDTD

model was simplified to have only 2 layers: the freshwater saturated layer, and the

interface layer. This simplification leads to easier understanding of the resulting data,

and for shorter computational time.


All of the data from the simulated receivers showed two arrivals: the direct

wave, and the interface reflection. To extract the interface reflection, a direct wave

with no

Figure 11 – The direct wave and interface reflection for a gradient of 132.67 mS/m2. The two
dimensional model is at 50 ns. The interface layer lies 1 meter below the source and the receiver, which
is 1.5 meters to the right of the source. Other properties are given in Table 1.
 
interface reflection was recorded (Figure 10). Then, the interface layer was added

(Figure 11), and the direct arrival was subtracted, resulting in the plot found in Figure

12. Notice that time >55 ns has not been plotted, as the “ringing” dominated the

record at late times.


Figure 12 – Plot of waves reflected off the freshwater/saltwater interface for various gradients. The
model is the same as in Figure 11. A greater gradient produces larger maximum amplitude reflections.

Figure 13 – Peak Amplitude vs. Conductivity Gradient of negative and positive components of
interface reflections in Figure 12.
 
For each run shown in Figure 12, the maxima and minima of the interface reflection

were recorded. A graph of the maximum amplitudes versus the gradient is shown in

Figure 13.

The relationship between the maximum amplitudes is not linear, but certainly not

entirely exponential: as the gradient increases, the change in increase of amplitude


becomes less. This phenomenon is more prevalent in the maximum negative

amplitudes.

The relationship between period of interface reflection and gradient is not a

clear, although it appears that the duration of the interface arrival generally decreases

as the gradient increases. A more interesting occurrence is the change in the shape of

the reflected wave. As the gradient increases, the location of the zero crossing

between the minimum and maximum shifts earlier in time. A comparison can be made

between different gradients by using the half width of the initial negative pulse. The

two times at which the amplitude is half the maximum amplitude are subtracted, and

the result is shown in Figure 14. As the gradient increases, the difference in time

decreases: the

Figure 14 – Plot of the gradient vs. the half-width of the negative initial pulse.

duration of the initial pulse decreases exponentially. This procedure can be used to

show that a difference in wave shape is found for varying gradients.


Another situation that was investigated was whether the angle at which the

direct pulse meets the interface layer varies the properties of the reflected wave.

Several

Figure 15 – Interface reflection for closely spaced receivers recording the same pulse. The interface
layer is 3 meters thick, and lies 2 meters below the source and the receivers, which are placed, starting
from the source location, every 2 meters. The distance from the source location for each plot is noted.
receivers were spaced every 2 meters and recorded the same pulse as it traveled back

from the interface layer. These receiver plots can be found in Figure 15: the plot does

not display the entire return because noise dominates the record. It can be seen from

Figure 15 that the angle of the incoming reflected wave does not significantly alter the

form of the arrival from receiver to receiver. Thus, an interface reflection does not

produce arrivals that distinctively change in character with increasing transmitter-

receiver spacing.

Conclusion

Because of the distinct ringing found during the runs, the FDTD model cannot

yet accurately represent the geometry of Key Largo, or a hypothesized interface layer.

However, several results were attainable using the improved algorithm coupled with

post-run spatial filtering. A change in the conductivity gradient at a

freshwater/saltwater interface affects both amplitude and the shape of the radar wave
reflected from the interface. If it can be assumed that a difference in 20% of the

amplitude is needed to recognize a change on the GPR record, a large change in the

conductivity gradient can be determined in the field: for example, a change in gradient

from 333 mS/m2 to 200 mS/m2 can be resolved, since the change in amplitude is

21%. If large changes in gradient such as this are found in the field, then ground

penetrating radar can be used with success.

 
 
 
 
 
 
References
Inman, J. A., S. Kruse, and H.L. Vacher. “Geophysical Imaging of the
Freshwater/Saltwater Interface on a Carbonate Island, Key Largo, Florida.” EOS,
Trans. AGU, 80, S121, 1999.
 
Reynolds, John M. An Introduction to Applied and Environmental Geophysics. John
Wiley & Sons, Chichester, West Sussex, England, 1997.

You might also like