You are on page 1of 16

WATER RESOURCES RESEARCH, VOL. 48, W08505, doi:10.

1029/2012WR011955, 2012

High-resolution water content estimation from surface-based


ground-penetrating radar reflection data by impedance inversion
C. Schmelzbach,1,2,3 J. Tronicke,1 and P. Dietrich4,5
Received 3 February 2012; revised 30 May 2012; accepted 7 June 2012; published 7 August 2012.
[1] Mapping hydrological parameter distributions in high resolution is essential to understand
and simulate groundwater flow and contaminant transport. Of particular interest is
surface-based ground-penetrating radar (GPR) reflection imaging in electrically resistive
sediments because of the expected close link between the subsurface water content and the
dielectric permittivity, which controls GPR wave velocity and reflectivity. Conventional
tools like common midpoint (CMP) velocity analysis provide physical parameter models of
limited resolution only. We present a novel reflection amplitude inversion workflow for
surface-based GPR data capable of resolving the subsurface dielectric permittivity and
related water content distribution with markedly improved resolution. Our scheme is an
adaptation of a seismic reflection impedance inversion scheme to surface-based GPR
data. Key is relative-amplitude-preserving data preconditioning including GPR
deconvolution, which results in traces with the source-wavelet distortions and propagation
effects largely removed. The subsequent inversion for the underlying dielectric permittivity
and water content structure is constrained by in situ dielectric permittivity data obtained
by direct-push logging. After demonstrating the potential of our novel scheme on a realistic
synthetic data set, we apply it to two 2-D 100 MHz GPR profiles acquired over a shallow
sedimentary aquifer resulting in water content images of the shallow (3–7 m depth) saturated
zone having decimeter resolution.
Citation: Schmelzbach, C., J. Tronicke, and P. Dietrich (2012), High-resolution water content estimation from
surface-based ground-penetrating radar reflection data by impedance inversion, Water Resour. Res., 48, W08505,
doi:10.1029/2012WR011955.

1. Introduction the close link between hydrologically relevant parameters


and the electrical properties governing GPR wave prop-
[2] An important issue in hydrological studies is the
agation [e.g., Huisman et al., 2003; van Overmeeren et al.,
detailed description of subsurface hydrological-parameter 1997]. In electrically resistive sediments such as sands,
distributions at the scales relevant to modeling groundwater
changes in the subsurface water content, qw, are widely rec-
flow and contaminant transport. Geophysical methods [e.g.,
ognized as the predominant cause of changes in the relative
Rubin and Hubbard, 2005] bridge the gap in resolution
dielectric permittivity, !r, which in turn controls GPR wave
and coverage between borehole logging and core analyses,
velocity and reflectivity [e.g., Kowalsky et al., 2001; Annan,
which provide very detailed but inherently 1-D information
2005; Knight and Endres, 2005]. In the saturated zone, qw is
[e.g., Kobr et al., 2005], and methods capturing large-scale
equivalent to porosity, f, which is a critical hydrological
average properties of an investigated area such as pumping
parameter related to permeability and storage, yet difficult to
and tracer tests [e.g., Leven and Dietrich, 2006]. The anal-
assess from borehole data.
ysis of efficiently acquired and noninvasive surface-based
[3] A common way to estimate the subsurface qw distri-
ground-penetrating radar (GPR) [e.g., Davis and Annan, 1989; bution from surface-based GPR data is common midpoint
Knight, 2001] data has received particular interest because of
(CMP) measurements, which require a multioffset acquisi-
tion strategy [e.g., Tillard and Dubois, 1995; Greaves et al.,
1
Institut für Erd- und Umweltwissenschaften, Universität Potsdam, 1996; Bradford, 2008]. Empirical relations such as Topp’s
Potsdam, Germany. equation [Topp et al., 1980] may be employed to relate !r
2
Department of Geosciences, Geophysics, Freie Universität Berlin, values derived from GPR velocity to qw. However, Bradford
Berlin, Germany. et al. [2009] pointed out that qw estimates from 100 MHz
3
Now at Institute of Geophysics, ETH Zurich, Zurich, Switzerland.
4
UFZ, Department Monitoring and Exploration Technologies, Helmholtz multioffset GPR data have a meter-scale lateral and vertical
Centre for Environmental Research, Leipzig, Germany. resolution when compared with spatially averaged neutron
5
Center of Applied Geoscience, University of Tübingen, Tübingen, porosity logs. This resolution capability, which is larger than
Germany. the dominant GPR wavelength, is comparable to the resolu-
Corresponding author: C. Schmelzbach, Institute of Geophysics, tion of cross-hole and vertical radar profiling (VRP) tra-
ETH Zurich, Sonneggstr. 5, CH-8092 Zurich, Switzerland. veltime tomography using 100 MHz [e.g., Tronicke et al.,
(cedric.schmelzbach@bluewin.de) 2002; Cassiani et al., 2004; Tronicke and Knoll, 2005;
©2012. American Geophysical Union. All Rights Reserved. Ernst et al., 2007a]. However, hydrological modeling may
0043-1397/12/2012WR011955 require higher-resolution parameter models [e.g., Hubbard
W08505 1 of 16
W08505 SCHMELZBACH ET AL.: HIGH-RESOLUTION WATER CONTENT FROM SURFACE GPR W08505

and Rubin, 2005]. GPR data interpretation techniques that reconstruct the large-scale physical parameter structure.
are able to provide very detailed physical property models Source signal and consequently data band limitation are also
such as amplitude-versus-offset (AVO) analysis [e.g., Baker, observed for GPR data. Whereas supplementary acoustic
1998; Zeng et al., 2000; Bradford and Deeds, 2006; Deparis impedance data are readily derived from sonic and density
and Garambois, 2009], finite-frequency tomography [e.g., borehole logs in the case of seismic data, corresponding tools
Buursink et al., 2008], and full-waveform tomography [e.g., to measure the electric impedance in situ in shallow uncon-
Ernst et al., 2007b; Gloaguen et al., 2007; Meles et al., 2010] solidated sediments at GPR operating frequencies have been
are generally much more demanding in terms of either data missing until lately. In this study, we revert to a recently
acquisition or computing time and, therefore, can often only developed direct-push (DP) tool to measure !r in situ. This
be used to construct models of limited spatial extent. DP tool is attached to the tip of a small diameter hollow steel
[4] The wavefield recorded by GPR systems is the result rod and pushed into the soft ground. Two ring electrodes with
of electromagnetic wave scattering from electric imped- an interelectrode spacing of !4 cm are used to measures
ance contrasts in the subsurface, largely analogous to !r with a frequency shift technique [Hilhorst, 1997] at an
seismic wave scattering at changes in the elastic properties. operating frequency of 20 MHz. Similar tools to measure !r
Seismic and GPR data analysis approaches are mutually in situ in soils have been presented by, for example, Shinn
transferable due to the strong mathematical analogies that et al. [1998], Vaz and Hopmans [2001], and Lin et al.
exist between seismic and electromagnetic wave propaga- [2006]. Studies linking DP !r log analysis to GPR data
tion [e.g., Carcione and Cavallini, 1995; Wapenaar, 2007; interpretation include, for example, Clement et al. [1997],
Belina et al., 2009b]. A testimony of the conceptual simi- Majer et al. [2001], and Schmelzbach et al. [2011b].
larities between standard constant-offset surface-based GPR [7] Here, we present a novel data interpretation approach
and poststack (zero-offset) seismic reflection data is the to invert the reflection amplitudes of surface-based GPR
fact that common GPR processing strategies are essentially measurements for high-resolution subsurface qw models with
identical to that established for seismic data [e.g., Neal, 2004; a vertical resolution in the quarter-wavelength range. The
Annan, 2005]. proposed workflow is an adaptation to GPR of the well-
[5] In seismic data analysis, workflows to invert poststack established model-based seismic impedance workflow. We
reflection data for the subsurface seismic impedance were begin by deriving the pertinent theoretical background of
first developed in the late 1970s and early 1980s [e.g., inverting the GPR reflection data to models of the parameters
Lavergne and Willm, 1977; Lindseth, 1979; Russell, 1988]. governing GPR wave propagation. Next, we illustrate and
In principle, seismic impedance inversion is capable of pro- assess the applicability of our new workflow on realistic
viding subsurface physical parameter models with a vertical synthetic GPR data. In particular, we demonstrate that the
resolution in the quarter-wavelength range (Rayleigh crite- simplifications regarding media properties and wave propa-
rion [Kallweit and Wood, 1982]). To the best of our knowl- gation underlying the inversion procedure are valid for GPR
edge, an analogous inversion of GPR reflection data has not data from typical sand/gravel aquifers. Finally, we apply our
yet been presented in the literature. One popular seismic new methodology to two 2-D field 100 MHz GPR data sets
reflection amplitude inversion technique is “model-based and derive a subsurface qw model with a decimeter resolution
impedance inversion.” This technique involves recovering for the shallow saturated zone of a sand-gravel aquifer.
the Earth’s reflectivity from suitably preconditioned data
under the assumption that the reflectivity represents a series 2. Methodology
of sparsely distributed spikes and constraining the inversion
by supplementary acoustic-impedance logging data [Russell, 2.1. Background
1988]. Data preconditioning, which involves relative- [8] In order to relate the recorded GPR signals to the
amplitude-preserving scaling, deconvolution, and migration, relevant subsurface physical properties, we assume high-
is critical to transform the recorded wavefield into data suit- frequency (10–1000 MHz) electromagnetic wave propagation
able for impedance inversion [see, e.g., Yilmaz, 2001]. These through nonmagnetic and low-loss media with frequency-
processing steps are readily adapted from seismic processing independent electric properties [see, e.g., Annan, 2005].
to GPR data analysis with the prominent exception of That is, the electric conductivity, s, is very small and the
deconvolution. Deconvolution is key to remove reflection medium’s magnetic permeability, m, is equal to that for free
image distortions due to the source wavelet and to improve space, m0. Typically, these assumptions are fulfilled for
the image accuracy. After deconvolution, the reflection data materials such as clay-free sand and gravel. Under these
are expected to more closely represent the subsurface r, ffiffidetermines
conditions, the relative dielectric permittivity, !p ffiffi
reflectivity than the unprocessed data. Because of the mixed- both the GPR propagation velocity (vGPR ¼ c0 = !r, where c0
phase GPR source wavelet, standard deconvolution routines is the speed of light in free space (!0.3 m/ns)) and the elec-
are often applied to GPR data with only limited success [see, trical impedance of a medium [e.g., Annan, 2005]
e.g., Belina et al., 2009a]. In contrast, the recently proposed
Z0
GPR deconvolution routine by Schmelzbach et al. [2011a] Z ¼ pffiffiffiffi ; ð1Þ
offers a robust way to estimate and remove the source !r
wavelet from GPR data.
[6] Because of the limited frequency bandwidth of seismic where Z0 is the impedance of free space (!377 W).
reflection data, supplementary data such as sonic logs or [9] Our inversion sequence is based on the assumption that
stacking velocities are necessary to constrain the impedance the subsurface can be represented as a stack of horizontal
inversion workflow in order to recover the full-bandwidth homogeneous layers. Furthermore, we assume that a suitably
subsurface impedance structure. In particular, the low- preconditioned GPR trace D(t) can be treated as a 1-D (plane
frequency information missing from seismic data is key to wave) wavefield that can be modeled as convolution of a

2 of 16
W08505 SCHMELZBACH ET AL.: HIGH-RESOLUTION WATER CONTENT FROM SURFACE GPR W08505

reflectivity series R(t) with a stationary band-limited wavelet [12] GPR deconvolution [Schmelzbach et al., 2011a] is
W(t), formulated as [e.g., Robinson and Treitel, 2000] key to remove the distortions of the recorded reflection data
due to the shape of the mixed-phase source wavelet. Standard
Dðt Þ ¼ Rðt Þ ⊗ W ðt Þ; ð2Þ deconvolution routines either involve the estimation of the
source wavelet from the data assuming it is minimum phase,
where ⊗ denotes the convolutional operator and t denotes which is often not met by GPR data, or require that the source
time. The corresponding normal incidence reflection coeffi- wavelet is known, which is very difficult for field data.
cient R at the boundary between the two media n and n + 1 is In contrast, the GPR deconvolution routine employed here
a function of the impedances of both media and is related via aims at retrieving the source wavelet from the data without
equation (1) to !r by [e.g., Ward and Hohmann, 2006] making assumptions about the phase spectrum of the wave-
pffiffiffiffiffi pffiffiffiffiffiffiffiffi let. Instead, GPR deconvolution is based on more restrictive
Znþ1 & Zn !rn & !rnþ1
Rn ¼ ¼ pffiffiffiffiffi pffiffiffiffiffiffiffiffi : ð3Þ assumptions about the amplitude value distribution of the
Zn þ Znþ1 !rn þ !rnþ1
deconvolution output than standard stochastic deconvolu-
tion. A further key feature is the use of prior information
The vector character of the electromagnetic waves and the
about the source wavelet, which is included in a Bayesian
source radiation patterns of the GPR antennas can be ignored
framework to constrain the wavelet estimation [see also
under the plane wave (far-field), normal incidence, and
Schmelzbach and Scherbaum, 2011]. We use GPR decon-
horizontal interfaces assumptions, which are implicit in
volution to estimate the original source wavelet for each trace
equations (2) and (3). The wavelet W(t) is assumed to be
and replace it by a symmetric, zero-phase W(t) with an
stationary, hence, the amplitude decay with propagation time
equalized amplitude spectrum over the signal bandwidth.
and dispersion [e.g., Annan, 1996; Irving and Knight, 2003;
Because this processing step is carried out trace by trace,
Bradford, 2007] are not modeled by equation (2). When
varying antenna coupling effects [Lampe and Holliger, 2003]
transformed to the frequency domain, equation (2) represents
are also largely removed. Not removing the source wavelet
the multiplication of the Fourier spectra of the broadband R(t)
from the GPR data would result in systematically erroneous
with the band-limited W(t), illustrating that D(t) is severely
final amplitude and time (depth) estimates of !r and q [e.g.,
band-limited compared to R(t) (e.g., a realistic 100 MHz GPR
Russell, 1988].
wavelet and resulting recorded data have a signal frequency
[13] Subsequent migration converts the recorded 2-D/3-D
range of less than two octaves i.e., !50–150 MHz).
wavefield into a 1-D wavefield with improved horizontal
2.2. Inversion strategy resolution [e.g., Stolt and Benson, 1986]. Given the assump-
[10] Our inversion scheme to extract subsurface physical tions underlying the convolutional model, we believe that
property models from the GPR reflection data consists of the seismic migration algorithms are sufficient for our purposes
following major steps: (1) deterministic amplitude scaling to [see, e.g., Fisher et al., 1992; Bradford, 2006]. After GPR
compensate for the reflection amplitude decay with propa- deconvolution and migration, the processed data D(t) are
gation time due to geometrical spreading, absorption and expected to meet the assumptions underlying equation (2).
transmission losses, (2) GPR deconvolution and migration to [14] In order to robustly extract an estimate of the band-
remove the GPR source wavelet, to produce an accurate limited reflectivity series 〈R(t)〉, where the angle brackets
image of the subsurface, and to increase the vertical and indicate band limitation, from the preconditioned data D(t),
horizontal resolution, (3) sparse-spike deconvolution to we use an L1 norm iterative reweighting least squares scheme
recover an estimate of the band-limited reflectivity from the for sparse-spike deconvolution [Sacchi, 1997]. Sparse-spike
preconditioned data, (4) band-limited integration and scal- deconvolution inverts equation (2) under the constraint that
ing to obtain an estimate of the band-limited impedance and the reflectivity series associated with a layered Earth model
the band-limited relative dielectric permittivity distribution, can be represented as a series of spikes [e.g., Oldenburg
(5) recovery of a full-bandwidth impedance estimate using et al., 1983]. This procedure provides the times and sizes of
supplementary data, and (6) estimation of the water content the largest reflectors and effectively recovers the lost high-
distribution. frequency part of the reflectivity.
[11] Reflection amplitudes decay with propagation time [15] The reflectivity estimates can be related to the sub-
due to geometrical spreading, absorption, and transmission surface impedance by reversing equation (3) to compute Zi+1
losses. The compensation for these effects is key for the later of layer i + 1 from the knowledge of Zi of the layer i
recovery of the correct reflectivity. We tested different through a recursive equation. Note that the errors accumulate
amplitude-scaling approaches (tn scaling; multiplication with in recursions. In generalized form and assuming that the
the inverse smoothed trace envelop [e.g., Gross et al., 2004]) impedance of the first layer Z1 is known, the impedance of
and found the most reliable amplitude recovery for both layer n is given by
synthetic and field data is using gain functions based on the n&1 "
Y #
1 þ Ri
inverse of the deterministic amplitude decay model given by Zn ¼ Z1 : ð5Þ
1 & Ri
" # i¼1
1
At ðt Þ ¼ A0 ðtÞ n expð&at Þ; ð4Þ
t Usually, an approximation to a continuous impedance dis-
tribution is used instead of equation (5), which is given by
where A0(t) and At(t) are the true and recorded amplitudes [e.g., Berteussen and Ursin, 1983; Oldenburg et al., 1983]
at time t, respectively [e.g., Russell, 1988]. The exponent n !
n&1
X
was set to 1 for the field data (3-D spherical spreading) and
Zn ≈ Z1 exp 2 Ri : ð6Þ
n = 0.5 was used to gain the 2-D synthetic data.
i¼1

3 of 16
W08505 SCHMELZBACH ET AL.: HIGH-RESOLUTION WATER CONTENT FROM SURFACE GPR W08505

Table 1. Processing Flow for GPR Impedance Inversion


Processing step Comment
1.GPR-related preprocessing e.g., time zero corrections, DC removal
2.Deterministic amplitude corrections Equation (4)
3.De-wow by high-pass filtering
4.Further signal-to-noise improvement e.g., suppression of shallow diffractions, first-arrival muting
5.GPR deconvolution to remove the mixed-phase GPR source wavelet Schmelzbach et al. [2011a]
6.Migration to transform the 2-D/3-D wavefield to 1-D
7.Optional correction for residual phase errors Levy and Oldenburg [1987]
8.Sparse-spike deconvolution to recover 〈R〉 Sacchi [1997]
9.Depth-to-time conversion of DP !r logs, correlation with GPR data to velocity information from CMP velocity analyses
cross-check the depth-to-time relation
10. Recovery of 〈Z〉 by band-limited integration and scaling; optionally Ferguson and Margrave [1996]; equation (8)
compute 〈!r〉
11. Recovery of Z by adding Zb equation (7); construction of Z b using, e.g., low-pass-filtered DP !r logs
12. Conversion of Z to !r and optionally to qw (or f) equation (1); e.g., Topp’s equation
13. Time-to-depth conversion of the physical parameter models for
interpretation

Equation (6) is valid for small reflection coefficients only decimeters can be expected for typical 100 MHz GPR data
(|R| < !0.3). In practice, we obtain an estimate of the in unsaturated sand-gravel environments (vGPR ≈ 0.1 m/ns).
band-limited impedance 〈Z(t)〉 by evaluating equation (6) in
the frequency domain by band-limited integration of 〈R(t)〉 3. Synthetic Example
[Ferguson and Margrave, 1996]. Because the correct scaling
of W(t) is usually not known when deconvolving D(t) (see [19] We now present a realistic synthetic example of how
equation (2)), also the correct scale of R(t) cannot be deter- our inversion workflow described above and summarized in
mined from the GPR data alone. Supplementary data such as, Table 1 can be used to derive high-resolution subsurface !r
for example, DP !r logs can be used to scale 〈R(t)〉 and 〈Z(t)〉 and qw images from surface-based GPR reflection amplitude
to the expected value range. data. A synthetic qw model was built consisting of horizontal
[16] Noise generally precludes the recovery of the lost layers with superimposed stochastic parameter variations
low-frequency part of R(t) from D(t) and the smooth back- (Figure 1a). The stochastic variations were defined by von
ground reflectivity and impedance trend, Zb(t), needs to be Kármán autocovariance functions [e.g., Goff and Jordan,
constructed from supplementary data [e.g., Russell, 1988]. 1988; Goff et al., 1994] with horizontal and vertical correla-
The full-bandwidth impedance is then obtained as follows tion lengths of 20 m and 1 m, respectively, leading to pre-
[e.g., Lindseth, 1979]: dominantly horizontal stratification. The mean and standard
deviation of the qw distribution are 0.22 and 0.019, respec-
Z ðtÞ ¼ hZ ðt Þi þ Z b ðt Þ; ð7Þ tively, with a minimum-to-maximum value range of 0.18 to
0.28, which we consider realistic for partially to fully water-
where Zb(t) may be derived from low-pass-filtered DP !r saturated sediments. In order to simulate the GPR wave
logs, high-quality CMP velocity functions, or geological propagation through our model, we converted the qw model
background information. Finally, the Z(t) estimates can be to subsurface !r values using Topp’s equation. For s, we
converted to !r(t) estimates using equation (1) and may be assumed a uniform value of 2 mS/m because conductivity
used to compute qw estimates using, for example, Topp’s variations with water content as predicted by, for example,
equation or site-specific models developed using available Archie’s law [Archie, 1942] would be minimal in our field
calibration data. Furthermore, the estimated !r(t) profiles may case. Throughout the simulation region, m was set to m0.
be converted to GPR propagation velocities and used for Figure 1b shows a 1-D !r profile extracted from the model at
time-to-depth conversion of the GPR reflection images. 10 m distance, which we included in the inversion as a sim-
[17] If no reliable Zb(t) can be derived, an estimate of the ulated synthetic DP !r log. One prominent feature of the
band-limited relative dielectric permittivity 〈!r(t)〉 may be subsurface model is a thin, !0.5 m thick high water content
obtained by evaluating the total differential of the rearranged (high !r or low velocity) layer at !4.5 m depth. The thickness
equation (1): of this layer should be compared with the dominant wave-
" # " # length of !0.9 m for an average !r of !12 and a dominant
Z0 2 hZ ðt Þi Z0 2 hZ ðt Þi frequency of 100 MHz. Converted to two-way travel times,
h!r ðtÞi ¼ &2 ≈ &2 ave ; ð8Þ
Z ðt Þ Z ðt Þ Z Z ave the thickness of this layer observed at !100 ns corresponds
to !15 ns (Figures 2a and 2b).
where Z ave represents a guessed average impedance value. At [20] We modeled a constant offset (1 m antenna separa-
sites where the large-scale GPR velocity structure is observed tion) GPR survey with 0.1 m trace spacing using GPRMAX2D,
to be almost constant, equation (8) may already provide inter- a 2-D finite difference time domain (FDTD) solution of
pretable models of relative !r and relative qw changes. Maxwell’s equations [Giannopoulos, 2005]. The source cur-
[18] In principle, the above outlined impedance inversion rent function was the first derivative of a Gaussian pulse with
scheme should be able to resolve the physical parameter a dominant frequency of 100 MHz (see Figure 3a for the
values of layers with a thickness greater than a quarter emitted source wavelet). After the numerical simulations,
wavelength (or a quarter of a period in time [e.g., Russell, we first compensated for the reflection amplitude decay with
1988]). Consequently, a resolution capability of a few propagation time using equation (4). The constant a was

4 of 16
W08505 SCHMELZBACH ET AL.: HIGH-RESOLUTION WATER CONTENT FROM SURFACE GPR W08505

Figure 1. (a) Input synthetic water content model. (b) Synthetic DP relative dielectric permittivity log that
would be measured at 10 m distance. (c) Synthetic 1 m common-offset 100 MHz GPR reflection image
obtained over the subsurface model shown in Figure 1a by FDTD modeling followed by scaling. (d) Syn-
thetic 1-D CMP velocity profile that would be measured at 10 m distance. (e) Final estimated water content
model.

5 of 16
W08505 SCHMELZBACH ET AL.: HIGH-RESOLUTION WATER CONTENT FROM SURFACE GPR W08505

Figure 2. Synthetic data traces extracted at 10 m distance for different processing stages (see also
Figure 1). (a) True (black line) and estimated (gray line) relative dielectric permittivity profile. (b) True
full-bandwidth reflectivity derived from Figure 2a. (c) GPR data trace extracted from the FDTD-modeled
data after scaling. (d) The GPR trace shown in Figure 2c after GPR deconvolution and migration. (e) Band-
limited reflectivity computed from Figure 2d. (f ) Band-limited impedance computed from Figure 2e.
(g) Low-frequency impedance background (gray line) and estimated full-bandwidth impedance (black
line). Traces in Figures 2c–2e were normalized to a standard deviation of 1.

estimated by fitting the data to reflectivity series estimated Convolution of the estimated source wavelet (gray wavelet in
from the DP !r log. Figure 1c shows the simulated GPR data Figure 3a) with the true reflectivity series (Figure 2b) yielded a
after scaling and muting the top parts to suppress waves trav- trace that closely matches the FDTD-modeled data trace
eling directly between the antennas. The poor correspondence (Trace 3 in Figure 3b). This observation confirms the reli-
between a GPR trace extracted from the scaled data at 10 m ability of the wavelet estimation and deconvolution output.
distance (Figure 2c) and the corresponding true reflectivity [21] We then migrated the deconvolved GPR image using
function (Figure 2b), which was derived by means of the phase shift algorithm of Gazdag [1978]. Figure 1d
equation (3) from the true !r trace shown in Figure 2a, illus- displays the simulated 1-D CMP velocity profile used for
trates the inherent distortions of the GPR data due to the source migration, which was derived by laterally averaging the
wavelet. For example, the maximum amplitudes of the distinct layered background model.
reflection at !150 ns is observed several nanoseconds later [22] Before subjecting the data to the actual impedance
than the corresponding spike in the reflectivity because of the inversion, we checked the data for a residual nonzero phase
delay of the maximum source wavelet amplitudes relative to wavelet using the technique proposed by Levy and Oldenburg
the onset of the wavelet (see Figure 3a). Because of the neg- [1987]. Residual phase distortions may be due to inaccuracies
ative sign of the maximum source wavelet amplitudes, of the GPR deconvolution or may have been introduced by
reflections in the GPR data trace show the opposite sign than other processing routines. Accurate impedance estimation
the corresponding spikes in the reflectivity. GPR deconvolu- requires the phase errors to be generally less than 10' [White,
tion [Schmelzbach et al., 2011a] was used to remove the 1997]. In the case of our synthetic data, we removed a min-
mixed-phase source wavelet (Figure 3a) and broaden the data imal 10' constant phase shift. Note that phase shifts <30'
amplitude spectrum within the signal bandwidth of !40–140 are not recognizable by eye. Then, we extracted 〈R(t)〉
MHz. Note the significantly improved similarity between the (Figure 2e) by sparse-spike deconvolution and performed the
deconvolved data (Figure 2d) and the reflectivity (Figure 2b). band-limited integration and scaling in the frequency domain
In order to assess the applicability of the convolutional model to estimate the band-limited impedance 〈Z(t)〉 (Figure 2f ).
to GPR deconvolution and to the subsequent impedance The full-bandwidth impedance Z(t) was obtained by adding
inversion workflow, we compared the FDTD modeled trace a 1-D impedance background model derived from the
shown in Figures 2c and 3b with a trace obtained by con- 0–40 MHz low-pass filtered DP !r log, which was hypo-
volving the known FDTD source wavelet (black wavelet in thetically collected in the center of the model. (Figure 2g).
Figure 3a) with the reflectivity series computed from the time- Finally, we converted the Z values to !r and qw values
converted !r model (Figure 2b). The close match between the employing equation (1) and Topp’s equation. Figure 1e dis-
FDTD-modeled trace and the convolution model trace con- plays the entire inverted qw model plotted as function of
firmed that convolution is well suited to simulate GPR data for depth.
low-resistive media [see, e.g., van Dam and Schlager, 2000].
6 of 16
W08505 SCHMELZBACH ET AL.: HIGH-RESOLUTION WATER CONTENT FROM SURFACE GPR W08505

distance is not representing the lateral changes sufficiently


at the intermediate-range scale, although the observed dis-
crepancies are small (<!10%). Conversely, the fine-scale
(>40 MHz) structures were well resolved confirming that
the proposed workflow, which is based on the simple 1-D
convolutional model (equation (2)), is applicable to realistic
GPR data.

4. Horstwalde Field Data Example


4.1. Database, Data Preconditioning, and Inversion
[24] We now discuss the application of our inversion
methodology to field GPR data acquired in a saturated
glaciofluvial environment at a site close to Horstwalde,
Germany [Niederleithinger, 2009]. Previous investigations
at this site identified a series of horizontally layered sand and
gravel units interrupted by thin organic material layers and
a groundwater table at !3 m depth [Linder et al., 2010;
Schmelzbach et al., 2011b; Tronicke et al., 2012]. For the two
2-D profiles reported here, we measured the GPR data in
continuous recording mode by moving a pair of 100 MHz
antennas with a constant separation of 1 m and !0.15 m trace
spacing using the surveying approach described by Böniger
and Tronicke [2010] (Figure 5a). To acquire the data, a
time sampling of 0.4 ns was employed and 16 vertical stacks
were taken at each point. The two transects were oriented
perpendicular to each other, with profile P1 oriented south–
north and profile P2 oriented west–east. An observed pene-
Figure 3. (a) True 100 MHz wavelet used for FDTD mod- tration depth of !15 m (!350 ns traveltime) and a generally
eling (black line) and wavelet estimated by the GPR decon- high signal-to-noise ratio indicated a geological environment
volution (light gray line). Wavelets were normalized to an
maximum absolute amplitude value of 1. (b) Trace 1, GPR
data trace extracted from the FDTD-modeled data after scal-
ing (same as Figure 2c); trace 2, convolution of the true
reflectivity series (Figure 2b) with the true 100 MHz wavelet
shown in Figure 3a; trace 3, convolution of the true reflec-
tivity series (Figure 2b) with the estimated 100 MHz wavelet
shown in Figure 3a. Traces in Figure 3b were normalized to a
standard deviation of 1. Each trace is plotted three times for
better visibility.

[23] Visual comparison of the true and estimated


qw models (Figures 1a and 1e) as well as the true and esti-
mated !r(t) traces (black and gray lines in Figure 2a) reveals
an overall close match between the two models. The overall
rms value of the differences is 0.017, corresponding to a
relative error of !7.5%. Note that the thin, subwavelength-
scale high qw layer at !4.5 m depth was well resolved,
illustrating the high-resolution capabilities of the proposed
workflow. In detail, the differences between the estimated
and true qw models (Figure 4a) increase away from the
hypothetical DP logging location at 10 m distance. For
example, the intermediate-scale qw structures at !2 m depth,
and 5–8 m and 12–15 m distance differ between the true and
inverted model by !0.02. When analyzing the amplitude
spectra of the trace-by-trace difference between the estimated
and true !r(t) shown in Figure 4b, we see that the discrepancy
between the two models is first of all concentrated in the Figure 4. (a) Difference between the estimated and true
<40 MHz range, whereas the differences are small and do water content model (Figures 1a and 1e). (b) Decadic loga-
not increase with distance from the logging location at 10 m rithm of the amplitude spectrum for each trace of the time
for the >40 MHz range. These observations illustrate that the domain difference between the estimated and true relative
0–40 MHz 1-D Zb model derived from one DP log at 10 m dielectric permittivity.
7 of 16
W08505 SCHMELZBACH ET AL.: HIGH-RESOLUTION WATER CONTENT FROM SURFACE GPR W08505

Figure 5. (a) Map of the Horstwalde survey area. P1 (black line) and P2 (gray line) mark the 2-D GPR
profile locations, and DP1–DP5 indicate the DP logging locations. (b) Original (dashed line) and edited
(solid line) CMP-based velocity profile at the DP1 location. (c) Raw DP !r log (dark gray line) and cali-
brated DP !r log (black line) matching the edited and !r-converted velocity profile (light gray line).

favorable for GPR imaging. Therefore, we expect low dis- high-amplitude coherent noise, which interfered with the
persive attenuation for this site. signals (see black arrow in Figure 6a). These events charac-
[25] An additional multioffset GPR profile collected along terized by steep apparent dips, were modeled in the f-k
the same line as P1 provided a 2-D CMP-based velocity domain and subtracted from the data (Figure 6b). Following
model that revealed a significant vertical velocity change the muting of the first arriving direct waves, GPR deconvo-
from !0.125 m/ns for the unsaturated zone to !0.075 m/ns lution was used to remove the mixed-phase GPR source
for the saturated zone across the groundwater table but wavelet from each trace (Figure 6c). Minor trace-to-trace
showed only mild lateral velocity variations of generally amplitude level variations after deconvolution were removed
<5%. Figure 5b displays a 1-D velocity profile extracted by rms balancing the 50–150 MHz band-pass-filtered traces.
from the 2-D velocity model at the crossing point of the two Phase shift migration using the 1-D velocity profile shown
GPR profiles. Because the groundwater table depth changed in Figure 5b repositioned dipping energy and collapsed
by !0.3 m between the acquisition of profiles P1/P2 and the diffractions, thereby transforming the 2-D data into a 1-D
surveying of the multioffset GPR data, which was carried out wavefield image with increased lateral resolution.
about half a year later, the depth of the velocity drop across [28] For the further impedance inversion, we concentrated
the groundwater table in the velocity profile was accordingly on the shallow saturated zone (i.e., !75–175 ns corre-
adjusted for the data analysis reported here. sponding to a depth interval of !3–7 m). The far-field
[26] DP !r logs were measured simultaneously with the assumption is likely violated in the shallow (<3 m) unsatu-
P1/P2 GPR data at five locations along the two profiles (i.e., rated zone and, although of minor importance at our site,
DP1–DP5; Figure 5a), with DP1 being located at the crossing dispersion effects may affect data at late (>175 ns) times.
point of the two profiles (for details, see Schmelzbach et al. After correcting for minor residual phase errors (step 7 in
[2011b]). A comparison of the DP logs with the GPR Table 1; Figure 6c), sparse-spike deconvolution of the pre-
velocity profiles converted to !r values revealed that the DP conditioned data provided estimates of 〈R(t)〉. Band-limited
tool consistently provided too low values (dark gray line in integration of 〈R(t)〉 and scaling to the impedance variation
Figure 5c). Later depth-to-time conversion tests and corre- range of the DP logs yielded 〈Z(t)〉 (Figure 7a).
lations with the GPR images, as well as laboratory tests [29] Matching the logging data, which were collected as a
(T. Vienken, personal communication, 2008) also indicated a function of depth, to the GPR images, which were recorded
calibration error of the DP tool. Therefore, we used collo- as a function of traveltime, required a high-quality velocity
cated velocity profiles and DP logs to derive a linear cali- function. The best match was found by scaling the original
bration function to correct all !r logs (black line in Figure 5c). CMP-based 1-D velocity function (dashed line in Figure 5b)
[27] Following the processing flow summarized in Table 1, by 105% and additionally shifting the logs by &6 ns thereby
the GPR data preconditioning involved subtracting DC shifts correcting for minor velocity-picking errors and GPR
from each trace, time-zero adjustments and corrections for velocity variations in the unsaturated zone (solid line in
minor (<1 m) topographic variations by time shifting, fol- Figure 5b; see also Figure 8).We then constructed the
lowed by deterministic amplitude corrections (equation (4)). impedance background model (Zb) by interpolation of the
“De-wow” filtering by high-pass frequency filtering (pass time-converted and low-pass-filtered (pass frequencies
frequencies >20 MHz) proved to preserve signals better than <40 MHz) DP logs (Figure 7b). Figure 7c displays the final
nonlinear residual median filtering [Gerlitz et al., 1993] model after combining 〈Z(t)〉 and Zb (equation (7); step 11
(Figure 6a). Diffractions originating from a system of pipes in Table 1).
buried at !1.5 m depth [Böniger and Tronicke, 2012] caused
8 of 16
W08505 SCHMELZBACH ET AL.: HIGH-RESOLUTION WATER CONTENT FROM SURFACE GPR W08505

Figure 6. (a) P1 data processed up to step 3 in Table 1. The blue arrows mark the reflection from the
groundwater table at !3 m depth, the black arrow indicates a high-amplitude diffraction from a shallow
pipe, and the dark gray box marks the portion of the data shown in Figures 6b and 6d. (b) P1 data portion
input to GPR deconvolution after diffraction removal (step 5 in Table 1). (c) Source wavelets estimated by
the GPR deconvolution. Every other wavelet is plotted as a wiggle trace. (d) P1 data portion input to sparse-
spike deconvolution (step 8 in Table 1). Two-way traveltime converted to depth is indicated on the right.
DP logging locations and crossing point of profiles P1 and P2 are indicated in Figures 6a, 6b, and 6d.

4.2. Assessment and Final Water Content Models amplitude scaling. Because the velocity of the unsaturated
[30] The close match between a 1-D profile extracted from zone (depth <3 m) was not resolved by our inversion
the estimated Z model and an impedance profile computed approach and due to the causal source wavelet used for the
from the DP1 !r log shown in Figure 9a highlights that the FDTD simulations compared to the zero-phase wavelet
impedance structure derived from the GPR reflection data remaining in the deconvolved field data, the simulated data
strongly correlates with the subsurface !r structure captured were shifted by &21.2 ns to allow a one-to-one comparison
by the DP sonde. Note that the DP logs were used in the within the target time window (75–175 ns). The overall close
impedance inversion workflow only to scale 〈Z(t)〉 and to correspondence between the field data and synthetic data is
construct Zb. The overall high correlation of P1 and P2 Z a strong indication that the impedance inversion resulted in
traces at their crossing point further confirms the repro- an !r image of the target zone that is reasonably close to
ducibility of the Z images (Figure 9b). As a further test, we the true subsurface structure. Trace-by-trace comparisons
repeated the impedance inversion of the P1 GPR data but (Figures 10c–10e) further highlight the close match in timing
excluded the DP1 data (i.e., using only DP2 and DP4 for and relative amplitudes between field and synthetic data. In
scaling and Zb construction) to assess the impact of the DP1 summary, the high correlation between DP logs and the final
log on the impedance results (Figure 9c). Whereas the large- inversion result (Figure 9), and the close match between the
scale trend of the recovered Z structure significantly differs field data and the FDTD simulations based on the inversion
from the final model, the fine-scale structure and, hence, the results (Figure 10) suggest that our inversion approach reli-
recovered relative !r correlates well with the DP logging ably resolved the subsurface !r structure.
information. [32] Figure 11 displays the final estimated qw images for
[31] As a final test, we input the estimated and depth- the P1 and P2 GPR data spliced together with the DP logging
converted !r structure (see Figure 7c) into the GPRMAX2D information for the southern part of our investigation area.
FDTD scheme to compute synthetic data with the same Topp’s equation was used to convert the !r values from both
acquisition parameters as the field data P1 profile. We used the inversion results and DP logs to qw. GPR wavelengths
a symmetric Ricker-type source wavelet with a frequency between 0.6 and 0.8 m result in a theoretical vertical resolu-
content comparable to the processed field data. Note that a tion of 0.15–0.20 m for the qw images. Overall, the DP logs
negative time shift of half the wavelet duration is sufficient to and the inverted qw profiles show a close match, notwith-
transform a causal Ricker wavelet into a zero-phase wavelet standing the fact that surface-based GPR data and DP logging
that is symmetric around time zero. Figure 10 displays the have different support volumes. General features of the
field data after deconvolution but before migration (step 5 in resolved qw models are low qw values for depths <4.5 m,
Table 1) together with the synthetic section after deterministic intermediate qw values for 4.5–6 m depth, and high qw values
for depths >6 m. The P1 and P2 data show a somewhat better
9 of 16
W08505 SCHMELZBACH ET AL.: HIGH-RESOLUTION WATER CONTENT FROM SURFACE GPR W08505

Figure 7. (a) Portion of the estimated band-limited impedance for profile P1. (b) Portion of the low-
frequency impedance background for profile P1. (c) Portion of the estimated full-bandwidth impedance
for profile P1. The color bar indicates also the corresponding full-bandwidth relative dielectric permittivity
values. DP logging locations and crossing point of profiles P1 and P2 are indicated in Figures 7a–7c.

10 of 16
W08505 SCHMELZBACH ET AL.: HIGH-RESOLUTION WATER CONTENT FROM SURFACE GPR W08505

and high qw layers that could be relevant for the local-scale


hydrological regime.

5. Discussion
[33] Our inversion approach is based on the assumption
that GPR plane wave propagation through low-loss media
with frequency-independent properties can be modeled by
the convolutional model (equation (2)). We believe that
the assumptions underlying this modeling approach do not
restrict the applicability of our workflow to typical hydro-
logical studies. The success of GPR imaging would most
likely only be limited at sites with significant dispersive
attenuation (e.g., clay-rich sediments [Arcone et al., 2008]).
A large number of successful GPR reflection imaging studies
document the wide range of environments suitable for GPR
imaging [e.g., Beres and Haeni, 1991; Smith and Jol, 1992;
Figure 8. (a) DP1 reflectivity log band-pass filtered to the Beres et al., 1995, 1999; McMechan et al., 1997; Tronicke
same frequency content as the field data and depth-to-time et al., 1999; Nitsche et al., 2002]. If moderate dispersion
converted using the original CMP velocity profile (dashed effects are observed, techniques such as proposed by Turner
line in Figure 5b). (b) Band-pass filtered DP1 reflectivity [1994] or Irving and Knight [2003] could be applied to
log shown in Figure 8a but depth-to-time converted using a ensure a stationary wavelet. In principle, our workflow is
105% scaled CMP velocity profile (solid line in Figure 5b). applicable to data from environments with low to moderate
(c) Band-pass-filtered DP1 reflectivity log shown in Figure 8b subsurface relief only, where the flat reflector assumption
with an additional time shift of &6 ns applied. (d) P1 trace holds locally. In case of shallowly dipping reflectors, modi-
extracted at the DP1 location from the deconvolved and fied migration routines may be used to correct for the antenna
migrated data (Figure 6c). In Figures 8a–8d, each trace is radiation pattern [e.g., Moran et al., 2000; Streich et al.,
plotted three times for better visibility. 2007; Streich and van der Kruk, 2007].
[34] Low-frequency information is key to recover the
match for depths >4.5 m compared to shallower depths, intermediate-scale to large-scale physical property distribu-
which may be due to minor differences in the background tion (see Figures 4 and 9), which is well known from seismic
model derived from DP1, DP2, and DP4 for P1 and DP1, impedance inversion [e.g., Russell, 1988]. Supplementary
DP3, and DP5 for P2. Regarding the fine-scale structures, the data such as DP !r logs, GPR velocity models from CMP
final qw models show a number of relatively thin (<1 m) low velocity analyses, or geological background information are

Figure 9. Impedance profiles extracted at the crossing point of profiles P1 and P2 (location of DP1).
(a) DP1 impedance log (black line), low-frequency impedance background (Zb, light gray line), and P1
impedance estimate (Z est, dark gray line). (b) P1 (dark gray line) and P2 (light gray line) impedance
estimate plotted together with the DP1 impedance log (black line). (c) P1 low-frequency impedance back-
ground (Zb, light gray line) and impedance estimate (Z est, dark gray line) derived without DP1 log data.
Arrows mark examples of close fit.
11 of 16
W08505 SCHMELZBACH ET AL.: HIGH-RESOLUTION WATER CONTENT FROM SURFACE GPR W08505

Figure 10. (a) P1 section after deconvolution (processed up to step 5 in Table 1). (b) FDTD-modeled
synthetic data based on the inverted relative dielectric permittivity model (Figure 7c). The synthetic data
were shifted by &21.2 ns to match the field data (see text for details). Sections in Figures 10a and 10b were
scaled to a rms amplitude value of 1. DP logging locations DP1 and DP2 and crossing point of profiles P1
and P2 are indicated in Figures 10a and 10b. (c–e) Traces extracted at 25 m, 31 m, and 36 m distance from
Figures 10a and 10b.

12 of 16
W08505

13 of 16
SCHMELZBACH ET AL.: HIGH-RESOLUTION WATER CONTENT FROM SURFACE GPR

Figure 11. (a and c) Water content estimates computed from the P1 and P2 data. Estimated water content profiles extracted at
the DP logging locations and converted DP water content logs are shown at the DP logging locations. (b) Map of the survey
area showing the locations of image planes A-DP1-A′ and B-DP1-B′ and DP logging locations.
W08505
W08505 SCHMELZBACH ET AL.: HIGH-RESOLUTION WATER CONTENT FROM SURFACE GPR W08505

necessary to substitute the missing low-frequency reflection Arcone, S., S. Grant, G. Boitnott, and B. Bostick (2008), Complex per-
information. Because GPR source wavelets often have dis- mittivity and clay mineralogy of grain-size fractions in a wet silt soil,
tinct band-limited frequency characteristics with a sharp Geophysics, 73(3), J1–J13, doi:10.1190/1.2890776.
Baker, G. S. (1998), Applying AVO analysis to GPR data, Geophys. Res.
amplitude spectrum, the problem of missing low-frequency Lett., 25(3), 397–400, doi:10.1029/97GL03773.
reflection information may be more severe for GPR data than Belina, F., B. Dafflon, J. Tronicke, and K. Holliger (2009a), Enhancing
for seismic data. Furthermore, the missing low-frequency the vertical resolution of surface georadar data, J. Appl. Geophys.,
reflection information may become a more severe problem 68(1), 26–35, doi:10.1016/j.jappgeo.2008.08.011.
Belina, F. A., J. R. Ernst, and K. Holliger (2009b), Inversion of crosshole
with increasing nominal GPR antenna frequency, probably seismic data in heterogeneous environments: Comparison of waveform
limiting the application of the workflow to data from low- and ray-based approaches, J. Appl. Geophys., 68(1), 85–94, doi:10.1016/
frequency GPR systems (≤!100 MHz). Additionally, low- j.jappgeo.2008.10.012.
frequency induction phenomena and system-dependent noise Beres, M., and F. P. Haeni (1991), Application of ground-penetrating-radar
methods in hydrogeologic studies, Ground Water, 29(3), 375–386,
(“wow”) usually contaminate the low end of the GPR signal doi:10.1111/j.1745-6584.1991.tb00528.x.
amplitude spectrum and “de-wow” filtering may also affect Beres, M., A. Green, P. Huggenberger, and H. Horstmeyer (1995), Map-
the critical low-frequency signal portion. ping the architecture of glaciofluvial sediments with three-dimensional
[35] The computationally most expensive step in our georadar, Geology, 23(12), 1087–1090.
workflow displayed in Table 1 is GPR deconvolution, which Beres, M., P. Huggenberger, A. G. Green, and H. Horstmeyer (1999), Using
two- and three-dimensional georadar methods to characterize glaciofluvial
involves estimating a mixed-phase source wavelet for each architecture, Sediment. Geol., 129(1–2), 1–24.
trace in several tens to few hundred iterations. All other Berteussen, K. A., and B. Ursin (1983), Approximate computation of the
processing steps except migration and sparse-spike decon- acoustic impedance from seismic data, Geophysics, 48(10), 1351–1358,
volution are single trace and one-iteration operations that doi:10.1190/1.1441415.
have a small computational load. Two-dimensional migra- Böniger, U., and J. Tronicke (2010), On the potential of kinematic GPR
surveying using a self-tracking total station: Evaluating system cross-talk
tion and iterative sparse-spike deconvolution of each trace and latency, IEEE Trans. Geosci. Remote Sens., 48(10), 3792–3798,
require a comparable computational effort, which is small to doi:10.1109/TGRS.2010.2048332.
intermediate in terms of computation time in relation to the Böniger, U., and J. Tronicke (2012), Subsurface utility extraction and
other processing steps. characterization: Combining GPR symmetry and polarization attributes,
IEEE Trans. Geosci. Remote Sens., 50(3), 736–746, doi:10.1109/
TGRS.2011.2163413.
6. Conclusions Bradford, J. H. (2006), Applying reflection tomography in the postmigra-
tion domain to multifold ground-penetrating radar data, Geophysics,
[36] We have presented a novel methodology for the 71(1), K1–K8, doi:10.1190/1.2159051.
estimation of qw (or f for fully water-saturated conditions) Bradford, J. H. (2007), Frequency-dependent attenuation analysis of
models from standard surface-based constant-offset GPR ground-penetrating radar data, Geophysics, 72(3), J7–J16, doi:10.1190/
reflection data using an impedance inversion workflow. Our 1.2710183.
Bradford, J. H. (2008), Measuring water content heterogeneity using multi-
methodology represents a significant step forward in the fold GPR with reflection tomography, Vadose Zone J., 7(1), 184–193,
GPR data interpretation in that the inversion of reflection doi:10.2136/vzj2006.0160.
amplitudes delivers quantitative physical parameter images Bradford, J. H., and J. C. Deeds (2006), Ground-penetrating radar theory
with a vertical resolution in the quarter-wavelength range, and application of thin-bed offset-dependent reflectivity, Geophysics,
71(3), K47–K57, doi:10.1190/1.2194524.
which is comparable to the resolution capabilities of Bradford, J. H., W. P. Clement, and W. Barrash (2009), Estimating porosity
full-waveform tomography. By contrast, the resolution with ground-penetrating radar reflection tomography: A controlled 3-D
of standard CMP-based velocity analyses and traveltime experiment at the Boise Hydrogeophysical Research Site, Water Resour.
tomography is significantly inferior. Key prerequisites of Res., 45, W00D26, doi:10.1029/2008WR006960.
our workflow are the application of a relative-amplitude- Buursink, M. L., T. C. Johnson, P. S. Routh, and M. D. Knoll (2008),
Crosshole radar velocity tomography with finite-frequency Fresnel vol-
preserving data preconditioning including GPR deconvolu- ume sensitivities, Geophys. J. Int., 172(1), 1–17, doi:10.1111/j.1365-
tion and DP !r logs or other supplementary data to constrain 246X.2007.03589.x.
the inversion workflow. The convolutional model and its Carcione, J. M., and F. Cavallini (1995), On the acoustic-electromagnetic
underlying assumptions (plane waves, horizontal interfaces, analogy, Wave Motion, 21(2), 149–162, doi:10.1016/0165-2125(94)
00047-9.
stationary source signal) are an integral part of our approach. Cassiani, G., C. Strobbia, and L. Gallotti (2004), Vertical radar profiles
We have discussed here the application of our workflow to for the characterization of deep vadose zones, Vadose Zone J., 3(4),
2-D GPR data, but the extension to 3-D is straightforward. 1093–1105, doi:10.2113/3.4.1093.
Clement, W. P., S. Cardimona, A. L. Endres, and K. Kadinsky-Cade
[37] Acknowledgments. This work was supported by the “Deutsche (1997), Site characterization at the groundwater remediation field labora-
Forschungsgemeinschaft” (DFG) under TR 512/3-1 and DI 833/6-1. We tory, Leading Edge, 16(11), 1617–1621, doi:10.1190/1.1437538.
thank Stewart Greenhalgh for a thorough in-house review of the manuscript. Davis, J. L., and A. P. Annan (1989), Ground-penetrating radar for high-
We are grateful to Editor John Selker, the Associate Editor, and three anon- resolution mapping of soil and rock stratigraphy, Geophys. Prospect.,
ymous reviewers for their insightful comments and suggestions that helped 37(5), 531–551.
to improve this paper. Deparis, J., and S. Garambois (2009), On the use of dispersive APVO GPR
curves for thin-bed properties estimation: Theory and application to frac-
ture characterization, Geophysics, 74(1), J1–J12, doi:10.1190/1.3008545.
References Ernst, J. R., A. G. Green, H. Maurer, and K. Holliger (2007a), Application
of a new 2D time-domain full-waveform inversion scheme to crosshole
Annan, A. P. (1996), Transmission dispersion and GPR, J. Environ. Eng. radar data, Geophysics, 72(5), J53–J64.
Geophys., 1(B), 125–136, doi:10.4133/JEEG1.B.125. Ernst, J. R., H. Maurer, A. G. Green, and K. Holliger (2007b),
Annan, A. P. (2005), Ground-penetrating radar, in Near-Surface Geophysics, Full-waveform inversion of crosshole radar data based on 2-D finite-
Invest. Geophys., vol. 13, edited by D. K. Butler, chap. 11, pp. 357–438, difference time-domain solutions of Maxwell’s equations, IEEE Trans.
Soc. of Explor. Geophys., Tulsa, Okla. Geosci. Remote Sens., 45(9), 2807–2828.
Archie, G. E. (1942), The electrical resistivity log as an aid in determining Ferguson, R. J., and G. F. Margrave (1996), A simple algorithm for band-
some reservoir characteristics, Trans. Am. Inst. Min. Metall. Pet. Eng., limited impedance inversion, CREWES Res. Rep., 8, 1–10.
146, 54–62.

14 of 16
W08505 SCHMELZBACH ET AL.: HIGH-RESOLUTION WATER CONTENT FROM SURFACE GPR W08505

Fisher, E., G. A. McMechan, A. P. Annan, and S. W. Cosway (1992), Linder, S., H. Paasche, J. Tronicke, E. Niederleithinger, and T. Vienken
Examples of reverse-time migration of single-channel, ground-penetrating (2010), Zonal cooperative inversion of crosshole P-wave, S-wave, and
radar profiles, Geophysics, 57(4), 577–586, doi:10.1190/1.1443271. georadar traveltime data sets, J. Appl. Geophys., 72(4), 254–262,
Gazdag, J. (1978), Wave equation migration with the phase-shift method, doi:10.1016/j.jappgeo.2010.10.003.
Geophysics, 43(7), 1342–1351, doi:10.1190/1.1440899. Lindseth, R. O. (1979), Synthetic sonic logs—A process for stratigraphic
Gerlitz, K., M. D. Knoll, G. M. Cross, R. D. Luzitano, and R. Knight interpretation, Geophysics, 44(1), 3–26, doi:10.1190/1.1440922.
(1993), Processing ground penetrating radar data to improve resolution Majer, E., K. H. Williams, J. E. Petterson, and T. E. Daley (2001), High
of near-surface targets, in Proceedings of the Symposium on the Appli- resolution imaging of vadose zone transport using crosswell radar and
cation of Geophysics to Environmental and Engineering Problems seismic methods, Tech. Rep. LBNL–49022, Lawrence Berkeley Natl.
(SAGEEP), vol. 6, pp. 561–574, Soc. of Explor. Geophys., Tulsa, Okla. Lab., Berkeley, Calif., doi:10.2172/792946.
Giannopoulos, A. (2005), Modelling ground penetrating radar by GprMax, McMechan, G. A., G. C. Gaynor, and R. B. Szerbiak (1997), Use of
Constr. Build. Mater., 19(10), 755–762, doi:10.1016/j.conbuildmat. ground-penetrating radar for 3-D sedimentological characterization of
2005.06.007. clastic reservoir analogs, Geophysics, 62(3), 786–796, doi:10.1190/
Gloaguen, E., B. Giroux, D. Marcotte, and R. Dimitrakopoulos (2007), 1.1444188.
Pseudo-full-waveform inversion of borehole GPR data using stochastic Meles, G., J. Van der Kruk, S. Greenhalgh, J. Ernst, H. Maurer, and
tomography, Geophysics, 72(5), J43–J51, doi:10.1190/1.2755929. A. Green (2010), A new vector waveform inversion algorithm for simul-
Goff, J., and T. Jordan (1988), Statistical modeling of seafloor morphology: taneous updating of conductivity and permittivity parameters from com-
Inversion of sea beam data for second-order statistics, J. Geophys. Res., bination crosshole/borehole-to-surface GPR data, IEEE Trans. Geosci.
93(B11), 13,589–13,608. Remote Sens., 48(9), 3391–3407, doi:10.1109/TGRS.2010.2046670.
Goff, J. A., K. Holliger, and A. Levander (1994), Modal fields: A new Moran, M. L., R. J. Greenfield, S. A. Arcone, and A. J. Delaney (2000),
method for characterization of random seismic velocity heterogeneity, Multidimensional GPR array processing using Kirchhoff migration,
Geophys. Res. Lett., 21(6), 493–496, doi:10.1029/94GL00311. J. Appl. Geophys., 43(2–4), 281–295, doi:10.1016/S0926-9851(99)
Greaves, R. J., D. P. Lesmes, J. M. Lee, and M. N. Toksöz (1996), Velocity 00065-8.
variations and water content estimated from multi-offset, ground-penetrating Neal, A. (2004), Ground-penetrating radar and its use in sedimentology:
radar, Geophysics, 61(3), 683–695, doi:10.1190/1.1443996. Principles, problems and progress, Earth Sci. Rev., 66(3–4), 261–330,
Gross, R., A. Green, H. Horstmeyer, and J. Begg (2004), Location and doi:10.1016/j.earscirev.2004.01.004.
geometry of the Wellington Fault (New Zealand) defined by detailed Niederleithinger, E. (2009), The BAM site for non-destructive testing meth-
three-dimensional georadar data, J. Geophys. Res., 109(B5), B05401, ods (NDT) in civil engineering, paper presented at International Founda-
doi:10.1029/2003JB002615. tion Congress and Equipment Expo 2009, Am. Soc. of Civil Eng.,
Hilhorst, M. A. (1997), Dielectric characterization of soil, PhD thesis, Orlando, Fla., 15–19 March.
Wageningen Agric. Univ., Wageningen, Netherlands. Nitsche, F. O., A. G. Green, H. Horstmeyer, and F. Büker (2002), Late
Hubbard, S. S., and Y. Rubin (2005), Introduction to hydrogeophysics, Quaternary depositional history of the Reuss Delta, Switzerland:
in Hydrogeophysics, edited by Y. Rubin and S. S. Hubbard, pp. 3–21, Constraints from high-resolution seismic reflection and georadar surveys,
Springer, Berlin. J. Quat. Sci., 17(2), 131–143, doi:10.1002/jqs.645.
Huisman, J. A., S. S. Hubbard, J. D. Redman, and A. P. Annan (2003), Oldenburg, D. W., T. Scheuer, and S. Levy (1983), Recovery of the
Measuring soil water content with ground penetrating radar, Vadose acoustic impedance from reflection seismograms, Geophysics, 48(10),
Zone J., 2(4), 476–491, doi:10.2136/vzj2003.4760. 1318–1337, doi:10.1190/1.1441413.
Irving, J. D., and R. J. Knight (2003), Removal of wavelet dispersion from Robinson, E. A., and S. Treitel (2000), Geophysical Signal Analysis, Soc.
ground-penetrating radar data, Geophysics, 68(3), 960–970, doi:10.1190/ of Explor. Geophys., Tulsa, Okla.
1.1581068. Rubin, Y., and S. S. Hubbard (Eds.) (2005), Hydrogeophysics, Water Sci.
Kallweit, R. S., and L. C. Wood (1982), The limits of resolution of zero- Technol. Libr., vol. 50, Springer, Berlin.
phase wavelets, Geophysics, 47(7), 1035–1046, doi:10.1190/1.1441367. Russell, B. H. (1988), Introduction to Seismic Inversion Methods, 177 pp.,
Knight, R. (2001), Ground penetrating radar for environmental applications, Soc. of Explor. Geophys., Tulsa, Okla., doi:10.1190/1.9781560802303.
Annu. Rev. Earth Planet. Sci., 29(1), 229–255, doi:10.1146/annurev. Sacchi, M. D. (1997), Reweighting strategies in seismic deconvolution,
earth.29.1.229. Geophys. J. Int., 129(3), 651–656, doi:10.1111/j.1365-246X.1997.
Knight, R. J., and A. L. Endres (2005), An introduction to rock physics tb04500.x.
principles for near-surface geophysics, in Near-Surface Geophysics, Schmelzbach, C., and F. Scherbaum (2011), Bayesian frequency-domain
Invest. Geophys., vol. 13, edited by D. K. Butler, chap. 3, pp. 31–70, mixed-phase wavelet estimation and deconvolution, paper presented
Soc. of Explor. Geophys., Tulsa, Okla. at 73rd EAGE Conference and Exhibition, Eur. Assoc. of Geosci. and
Kobr, M., S. Mareš, and F. Paillet (2005), Geophysical well logging: Bore- Eng., Vienna.
hole geophysics for hydrogeological studies: Principles and applications, Schmelzbach, C., F. Scherbaum, J. Tronicke, and P. Dietrich (2011a),
in Hydrogeophysics, edited by Y. Rubin and S. S. Hubbard, pp. 291–331, Bayesian frequency-domain blind deconvolution of ground-penetrating
Springer, Berlin. radar data, J. Appl. Geophys., 75(4), 615–630, doi:10.1016/j.jappgeo.
Kowalsky, M. B., P. Dietrich, G. Teutsch, and Y. Rubin (2001), Forward 2011.08.010.
modeling of ground-penetrating radar data using digitized outcrop images Schmelzbach, C., J. Tronicke, and P. Dietrich (2011b), Three-dimensional
and multiple scenarios of water saturation, Water Resour. Res., 37(6), hydrostratigraphic models from ground-penetrating radar and direct-push
1615–1625, doi:10.1029/2001WR900015. data, J. Hydrol., 398, 235–245, doi:10.1016/j.jhydrol.2010.12.023.
Lampe, B., and K. Holliger (2003), Effects of fractal fluctuations in Shinn, J. D., D. A. Timian, R. M. Morey, G. Mitchell, C. L. Antle, and
topographic relief, permittivity and conductivity on ground-penetrating R. Hull (1998), Development of a CPT deployed probe for in situ mea-
radar antenna radiation, Geophysics, 68(6), 1934–1944, doi:10.1190/ surement of volumetric soil moisture content and electrical resistivity,
1.1635047. Field Anal. Chem. Technol., 2(2), 103–109, doi:10.1002/(SICI)1520-
Lavergne, M., and C. Willm (1977), Inversion of seismograms and pseudo 6521(1998)2:2<103::AID-FACT6>3.0.CO;2-X.
velocity logs, Geophys. Prospect., 25(2), 231–250, doi:10.1111/j.1365- Smith, D. G., and H. M. Jol (1992), Ground-penetrating radar investigation
2478.1977.tb01165.x. of a Lake Bonneville Delta, Provo level, Brigham City, Utah, Geology,
Leven, C., and P. Dietrich (2006), What information can we get from 20(12), 1083–1086.
pumping tests? Comparing pumping test configurations using sensitivity Stolt, R. H., and A. Benson (1986), Seismic Migration: Theory and
coefficients, J. Hydrol., 319(1–4), 199–215, doi:10.1016/j.jhydrol.2005. Practice, 382 pp., Geophys. Press, London.
06.030. Streich, R., and J. van der Kruk (2007), Accurate imaging of multicom-
Levy, S., and D. W. Oldenburg (1987), Automatic phase correction of ponent GPR data based on exact radiation patterns, IEEE Trans. Geosci.
common-midpoint stacked data, Geophysics, 52(1), 51–59, doi:10.1190/ Remote Sens., 45(1), 93–103, doi:10.1109/TGRS.2006.883459.
1.1442240. Streich, R., J. van der Kruk, and A. G. Green (2007), Vector migration
Lin, C. P., S. H. Tang, and C. C. Chung (2006), Development of of standard copolarized 3D GPR data, Geophysics, 72(5), J65–J75,
TDR penetrometer through theoretical and laboratory investigations: doi:10.1190/1.2766466.
1. Measurement of soil dielectric permittivity, Geotech. Test. J., 29(4), Tillard, S., and J.-C. Dubois (1995), Analysis of GPR data: Wave prop-
306–313, doi:10.1520/GTJ14093. agation velocity determination, J. Appl. Geophys., 33(1–3), 77–91,
doi:10.1016/0926-9851(95)90031-4.

15 of 16
W08505 SCHMELZBACH ET AL.: HIGH-RESOLUTION WATER CONTENT FROM SURFACE GPR W08505

Topp, G. C., J. L. Davis, and A. P. Annan (1980), Electromagnetic determi- van Overmeeren, R. A., S. V. Sariowan, and J. C. Gehrels (1997), Ground
nation of soil water content: Measurements in coaxial transmission lines, penetrating radar for determining volumetric soil water content; results
Water Resour. Res., 16(3), 574–582. of comparative measurements at two test sites, J. Hydrol., 197(1–4),
Tronicke, J., and M. D. Knoll (2005), Vertical radar profiling: Influence 316–338, doi:10.1016/S0022-1694(96)03244-1.
of survey geometry on first-arrival traveltimes and amplitudes, J. Appl. Vaz, C. M. P., and J. W. Hopmans (2001), Simultaneous measurement of soil
Geophys., 57(3), 179–191. penetration resistance and water content with a combined penetrometer-
Tronicke, J., N. Blindow, R. Gross, and M. A. Lange (1999), Joint applica- TDR moisture probe, Soil Sci. Soc. Am. J., 65(1), 4–12, doi:10.2136/
tion of surface electrical resistivity- and GPR-measurements for ground- sssaj2001.6514.
water exploration on the island of Spiekeroog—Northern Germany, Wapenaar, K. (2007), General representations for wavefield modeling and
J. Hydrol., 223(1–2), 44–53, doi:10.1016/S0022-1694(99)00111-0. inversion in geophysics, Geophysics, 72(5), SM5–SM17, doi:10.1190/
Tronicke, J., P. Dietrich, U. Wahlig, and E. Appel (2002), Integrating sur- 1.2750646.
face georadar and crosshole radar tomography: A validation experiment Ward, S. H., and G. W. Hohmann (2006), Electromagnetic theory for
in braided stream deposits, Geophysics, 67(5), 1516–1523, doi:10.1190/ geophysical applications, in Electromagnetic Methods in Applied Geo-
1.1512747. physics, Invest. Geophys., vol. 3, edited by M. N. Nabighian, chap. 4,
Tronicke, J., H. Paasche, and U. Böniger (2012), Crosshole traveltime pp. 131–311, Soc. of Explor. Geophys., Tulsa, Okla.
tomography using particle swarm optimization: A near-surface field White, R. (1997), The accuracy of well ties: Practical procedures and exam-
example, Geophysics, 77(1), R19–R32, doi:10.1190/geo2010-0411.1. ples, paper presented at 67th SEG Annual Meeting and International
Turner, G. (1994), Subsurface radar propagation deconvolution, Geophysics, Exposition, Soc. of Explor. Geophys., Dallas, Tex.
59(2), 215–223, doi:10.1190/1.1443583. Yilmaz, O. (2001), Seismic Data Analysis: Processing, Inversion, and
van Dam, R. L., and W. Schlager (2000), Identifying causes of ground- Interpretation of Seismic Data, Soc. of Explor. Geophys., Tulsa, Okla.
penetrating radar reflections using time-domain reflectometry and sedi- Zeng, X., G. A. McMechan, and T. Xu (2000), Synthesis of amplitude-
mentological analyses, Sedimentology, 47(2), 435–449, doi:10.1046/ versus-offset variations in ground-penetrating radar data, Geophysics,
j.1365-3091.2000.00304.x. 65(1), 113–125, doi:10.1190/1.1444702.

16 of 16

You might also like