You are on page 1of 11

736 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 50, NO.

3, MARCH 2012

Subsurface Utility Extraction and Characterization:


Combining GPR Symmetry and
Polarization Attributes
Urs Böniger, Member, IEEE, and Jens Tronicke

Abstract—Polarization of the electromagnetic wavefield has can be used to further characterize buried targets (see, e.g.,
significant implications for the acquisition and interpretation of [3]–[5]). Efficient detection and characterization of subsurface
ground-penetrating radar (GPR) data. Based on the geometrical utilities is of high importance considering the annual costs
and physical properties of the subsurface scatterer and the phys-
ical properties of its surrounding material, strong polarization arising from damage to unknown subsurface utilities during
phenomena might occur. Here, we develop an attribute-based construction or maintenance activities [6]. The 3-D GPR and its
analysis approach to extract and characterize buried utility pipes application for subsurface utility detection thus represent a field
using two broadside antenna configurations. First, we enhance of continuously growing importance. This study introduces
and extract the utilities by making use of their distinct sym- a signal- and image-processing-based approach to enhance
metric nature through the application of a symmetry-enhancing
image-processing algorithm known as phase symmetry. Second, and extract subsurface utilities, including their polarization
we assess the polarization characteristics by calculating two at- characteristics.
tributes (polarization angle and linearity) using principal compo- GPR data are commonly acquired using a perpendicular
nent analysis. Combination of attributes derived from these steps broadside antenna configuration, i.e., transmitter and receiver
into a novel depolarization attribute allows one to efficiently detect antennas are oriented parallel to each other and perpendicular
and distinguish different utilities present within 3-D GPR data.
The performance of our analysis approach is illustrated using to the direction of data acquisition (profile direction). As most
synthetic examples and evaluated using field examples (including GPR systems employ linearly polarized dipole antennas, the
a dual-configuration 3-D data set) collected across a field site, transmitting antenna emits an electromagnetic (EM) wavefield
where detailed ground-truth information is available. Our results whose electric field is polarized parallel to the long axis of the
demonstrate that the proposed approach allows for a more de- dipole, and the receiving antenna records only the component
tailed extraction and combination of utility relevant information
compared to approaches relying on single-component data and, parallel to its long axis. However, it has been noted that various
thus, eases the interpretation of multicomponent GPR data sets. targets of GPR surveys, such as buried pipes and fractures, have
polarization-dependent scattering characteristics (see, e.g., [4],
Index Terms—Multicomponent ground-penetrating radar
(GPR) data, phase symmetry, pipe detection, polarization [5], [7], and [8]). This implies that the visibility of a subsurface
analysis, principal component analysis (PCA). scatterer in the acquired data depends on the used antenna
configuration (e.g., broadside configuration) and its orientation
I. I NTRODUCTION with respect to the feature to be imaged. As a consequence,
certain subsurface objects might not be imaged using a single-

I N A VARIETY of archaeological, environmental, engi-


neering, and geological problems, ground-penetrating radar
(GPR) has become a popular geophysical technique to study
component perpendicular broadside antenna configuration. For
example, Radzevicius and Daniels [4] have found that the
backscattered fields from buried cylinders (such as pipes and
the shallow subsurface, mainly owing to its high-resolution cables) may be strongly depolarized depending on the orien-
imaging capabilities (see, e.g., [1] and [2]). Usually, the inter- tation of the cylinder relative to the antennas, the electrical
pretation of 2-D and 3-D GPR data sets is limited to a pure properties of the cylinders and the background material, as well
structural interpretation by, for example, tracking coherent re- as the depth and radius-to-wavelength ratio of the cylinders.
flection events from subsurface horizons or picking diffraction Thus, polarization-dependent scattering properties have impor-
hyperbolas caused by isolated objects and other discontinuities. tant implications for target detection, survey design, and data
Such a strategy focusing on target detection and depth deter- interpretation.
mination may be sufficient for some applications; however, it To address GPR polarization phenomena, various stud-
largely ignores the vectorial nature of the GPR wavefield, which ies examined the use of recording GPR data at multiple
transmitter–receiver configurations where, compared to the
Manuscript received April 19, 2010; revised September 29, 2010 and standard perpendicular broadside configuration, one or both
April 29, 2011; accepted July 10, 2011. Date of publication September 22,
2011; date of current version February 24, 2012. This work was supported by antennas were rotated to acquire different components of the
the Deutsche Forschungsgemeinschaft (DFG) under Project TR512/1-2. electric field. For example, Lehmann et al. [9] suggested the
The authors are with the Institut für Erd- und Umweltwissenschaften, combination of two GPR data sets recorded using perpen-
Universität Potsdam, 14476 Potsdam, Germany (e-mail: urs@geo.
uni-potsdam.de). dicular and parallel broadside antenna configurations to cre-
Digital Object Identifier 10.1109/TGRS.2011.2163413 ate a pseudoscalar wavefield. In a parallel broadside antenna

0196-2892/$26.00 © 2011 IEEE


BÖNIGER AND TRONICKE: SUBSURFACE UTILITY EXTRACTION AND CHARACTERIZATION 737

configuration, transmitter and receiver are oriented parallel to Research and Testing and is located ∼30 km south of Berlin,
each other and parallel to the direction of data acquisition. Germany. The geology at the site is predominantly charac-
Furthermore, advanced 3-D multiconfiguration (also known terized by glacial sands and gravels deposited during the last
as multicomponent) imaging approaches have been developed glaciation periods. During the installation of this test site,
in [10] and [11]. Gestel and Stoffa [12] made use of Alford consisting of several shallow-to-intermediate depth boreholes,
rotations to extract the main axis of subsurface anisotropy and numerous geophysical surveys have been conducted. The pri-
the orientation of buried objects from multipolarization 2-D mary goal of these surveys was to identify optimum borehole
GPR data, while the study in [13] has applied a similar approach locations which included the detection of subsurface utilities
to find the strike of fractures. Tsoflias et al. [5] made use of the installed during the past century when the site was used as
phase characteristics from 2-D multipolarization measurements a military training area. During these studies, we found that
to detect vertical fractures. In addition, Roberts and Daniels [3] the GPR technique provides the best performance at this site
have investigated GPR polarization phenomena associated with (compared to other geophysical techniques such as magnetic
typical GPR targets and found that polarization characteristics or EM induction mapping) and that multipolarization data sets
may be extremely useful for detecting and characterizing small- may further help to characterize the detected buried utilities.
diameter (relative to the incident wavelength) long circular
cylinders such as buried pipes and cables. Radzevicius and
B. Data Acquisition and Standard Processing
Daniels [4] have further investigated polarization and cylinder
scattering concepts relevant for GPR. Considering the response We conducted a dual-configuration 3-D GPR data set cover-
of metal and plastic pipes with varying diameters, these authors ing an area of ∼20 × 20 m and one 2-D profile perpendicular
also provide guidelines for survey design and data interpre- to three selected utility pipes using different antenna config-
tation. Nevertheless, multicomponent GPR surveying is far urations to further study polarization phenomena at our site.
from being a standard exploration tool, and further work is The data sets were acquired using cart-mounted unshielded
definitively needed for processing, analyzing, and interpreting 200-MHz linearly polarized dipole antennas using a fixed offset
such data sets. of 0.9 m. Positioning of the GPR traces was performed kinemat-
In this study, we present a dual-configuration 3-D GPR ically using a self-tracking total station (TTS). Details of this
processing flow using perpendicular and parallel broadside TTS-based acquisition approach are presented in [15] and [16].
antenna configurations to efficiently image and characterize The data sets were recorded using a time window of 300 ns, a
buried utility pipes. Our attribute-based processing and analysis sample interval of 0.2 ns, 16 vertical stacks, in- and crossline
strategy allows for improved target extraction and characteriza- trace spacings of ∼0.05 and ∼0.1 m, respectively, for the 3-D
tion by combining the geometrical and physical characteristics data sets, and a ∼0.05-m trace spacing for the 2-D profile.
from both components into a novel depolarization attribute. Our Processing the 3-D GPR data involved latency correction
strategy is illustrated using data collected at a field site where [16], zero-time correction, wow removal, frequency filtering,
detailed ground-truth information is available. In the following, and amplitude scaling using a smoothed average envelope
after introducing our field site and providing details regarding [1], [17]. It should be noted that the same amplitude scaling
data acquisition and standard processing, we illustrate and function was used for both data sets X11 and X22 recorded
discuss the main structural and polarization characteristics of using the perpendicular and parallel broadside antenna con-
our data. Then, we show how the structural information from figurations, respectively. The data were then gridded onto a
both data sets can be significantly enhanced by using a data rectangular grid (0.1 × 0.1 m) using a natural neighbor-based
attribute known as phase symmetry [14]. To characterize the algorithm [15], [18]–[20]. Afterward, the preprocessed data
polarization characteristics of our two-component data set, we cubes were migrated using a 3-D zero-offset Stolt algorithm
further employ two polarization attributes based on principal with a migration velocity of 0.111 m/ns derived from several
component analysis (PCA). Finally, through combination of the common-midpoint measurements. Fig. 1 shows time slices
proposed structural and physical attributes, we illustrate how through the unmigrated and migrated data cubes of the two
the major information from dual-configuration data sets can copolarized 3-D data sets X11 and X22 . Here, we focus on
be fused into one attribute which significantly enhances the the prominent linear features representing buried utility pipes.
interpretability of such data sets. When comparing the X11 and the X22 slices, polarization
effects are evident in the unmigrated and migrated data. Most
of these features show a dominant polarization perpendicular to
II. BACKGROUND
the orientation of the object. When comparing the unmigrated
In the following sections, we introduce the survey site and [Fig. 1(a)–(d)] and migrated data [Fig. 1(e)–(h)], we see that
our standard processing flow applied to the different data sets migration successfully focuses the diffracted energy and, thus,
presented in this study. significantly improves the resolution capabilities of the data.
To further study the polarization phenomena visible in the 3-D
data sets, we acquired a multicomponent 2-D profile by suc-
A. Survey Site
cessively rotating a broadside antenna configuration from 0◦
GPR data have been acquired at a test site located in to 180◦ using increments of 10◦ . Fig. 2 shows the survey
Horstwalde, Germany. This site has been installed by the geometry of this 2-D experiment, including approximative pipe
University of Potsdam and the Federal Institute for Materials geometries derived from Fig. 1. Only basic processing (latency
738 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 50, NO. 3, MARCH 2012

Fig. 1. (a)–(d) Unmigrated and (e)–(h) migrated time slices through the 3-D GPR data cubes at a two-way travel time of 21 ns (a, b, e, and f) and 25 ns (c, d, g,
and h). The symbols in the upper right corners indicate antenna orientation with respect to the in-line direction.

they are significantly attenuated at ∼ 90◦ . In contrast to Pipes


1 and 3, Pipe 2 is visible at all angles with maximum am-
plitudes around 90◦ antenna orientation. Summarizing Figs. 1
and 3, our data suggest that the survey site consists of two
types of pipes characterized by differences in reflection and
polarization strength. Based on these observations, we intro-
duce an attribute-based approach to extract and emphasize pipe
structures and potential polarization effects by performing the
following: 1) enhancing the geometry of the target features
(based on their characteristic symmetry) and 2) characterizing
the dominant polarization effects using PCA.

III. G EOMETRICAL E XTRACTION U SING


P HASE S YMMETRY
Interpretation of GPR data is commonly based on visual
Fig. 2. Illustration of the survey geometry of the 2-D multicomponent GPR
inspection of coherent events, which from our point of view is
experiment across three distinct subsurface features (utility pipes) to high- often based on the perception of feature symmetry. Symmetry
light the polarization phenomena using a broadside antenna configuration. represents an important biological characteristic in order to
The location of the pipes (labeled as 1, 2, 3, and 4) is sketched based on
Fig. 1. The dashed black line and the arrow indicate the profile location and
distinct between, e.g., animals and static environmental back-
direction, respectively. The sketch in the lower left corner illustrates the antenna grounds [21], [22], whereby its visual perception developed
configuration and the clockwise antenna rotation with respect to the profile during evolution. Man-made objects often exhibit intrinsic sym-
orientation.
metry (e.g., mirror or rotational) for different reasons, for ex-
correction, zero-time correction, wow removal, and amplitude ample, improved structural stability from an engineering point
scaling followed by gridding onto a trace spacing of 0.02 m) of view. Subsurface utilities also exhibit symmetry usually
was applied to the 2-D data in order to highlight the ob- facilitating visual perception when inspecting GPR time slices.
served polarization phenomena in minimally processed data. In order to detect and extract subsurface utilities from 3-D
Fig. 3 shows the resulting profiles obtained from this experi- GPR data in heterogeneous materials, we will exploit feature
ment. The three target events located around ∼10, ∼12.5, and symmetry.
∼14 m northing show interference phenomena at ∼11, ∼12, and Another observation based on visual perception is the signif-
∼13 m northing, and the corresponding subsurface scatterers icance of phase information for feature detection. References
are obviously located at different depths. In addition, Pipes 1 [23] and [24] showed how phase information contributes to the
and 3 show maximum amplitudes around 0◦ and 180◦ antenna human perception of discontinuities (such as edges) in images.
orientations (i.e., parallel broadside configuration). While the Following these studies, Kovesi [25], [26] introduced a phase-
corresponding diffraction hyperbolas are visible up to ∼ 60◦ , based feature extraction algorithm making use of 2-D rotational
BÖNIGER AND TRONICKE: SUBSURFACE UTILITY EXTRACTION AND CHARACTERIZATION 739

Fig. 3. Gridded profiles from our 2-D multicomponent GPR experiment (for survey geometry and angle definition, see Fig. 2). In the lower right corner, the
diffraction hyperbolas of the three pipes are sketched to highlight the target events. All the profiles are plotted using equal amplitude scaling.

log-Gabor filters [27] for edge and symmetry detection. Gabor corresponds to the angular difference to the specified filter
filters and, subsequently, log-Gabor filters have often been at- orientation, and σθ defines the standard deviation of the applied
tributed to have filtering properties similar to the ones found in Gaussian filter. Applied over multiple scales and orientations,
mammalian cortical cells [27]. Log-Gabor functions are defined one obtains a filter bank representation of the image (e.g., a
by a Gaussian amplitude spectrum (if plotted on a logarithmic GPR time slice) that consists of even- and odd-symmetric filter
frequency scale with a zero DC component) and provide a high outputs. As demonstrated for a 1-D signal by Kovesi [25],
spatial and frequency localization. Phase symmetry, a contrast- points with high symmetry are characterized by high magni-
invariant measure, was introduced by Kovesi [25] to highlight tudes in the even-symmetric filter output but low magnitudes in
feature symmetry within 2-D images. Russell et al. [28] just the odd-symmetric filter output. Therefore, phase symmetry at
recently applied a related phase-based edge detection algorithm the spatial coordinates i and j is defined as
to highlight faults in 3-D seismic data.
In the following, we briefly introduce our phase-symmetry-  
r,n ([|er,n (i, j)| − |or,n (i, j)|] − T )
based processing flow using a synthetic example in accordance S(i, j) = max  ,0
with our field data. Subsequently, we apply our processing flow r,n Ar,n (i, j) + 
to the dual-component 3-D GPR field data set. For a detailed (2)
introduction on phase-based edge and symmetry detection, we where the max operator limits the attribute to symmetric fea-
refer the interested reader to [25]. tures and er,n and or,n are the even- and odd-symmetric filter
outputs, respectively, at orientation r and scale n. T represents
A. Phase Symmetry a noise compensation term, and  was introduced to avoid
instabilities by zero division. Ar,n represents the magnitude per
As introduced by Kovesi [25], phase symmetry consists of orientation and scale and is defined as
2-D log-Gabor filter banks, obtained by elementwise multi-
plication of the radial component, i.e., the log-Gabor transfer 
function with an angular Gaussian scaling function Ar,n (i, j) = e2r,n (i, j) + o2r,n (i, j). (3)
 
− log(ω/ω0 )2 dθ2
G(ω, θ) = exp − (1)
2 · log(k/ω0 )2 2 · σθ2 The advantages of this phase-based symmetry measure are its
normalization to a range between 0 and 1 and its robustness
where ω and ω0 are the frequency range and the center fre- against changes in contrast as often observed in GPR time
quency, respectively. k corresponds to a scaling factor, dθ slice data.
740 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 50, NO. 3, MARCH 2012

Fig. 4. Suggested feature extraction approach applied to a synthetic image (Xm ) representing a single GPR time slice generated from utilities (Xu ) and
a geological background (Xg ). S(Xm ) is the result when phase symmetry is calculated directly from Xm . In order to improve the phase symmetry result,
preprocessing using Laplacian-based highboosting (L(Xm )) is applied, which produces a favorable extraction result ((S ◦ L)(Xm )). For details, see text.

B. Synthetic Examples symmetry image (S ◦ L)(Xm ) shows an increasing amount


of features from Xg because Laplacian-based highboosting
Fig. 4 shows our proposed feature extraction approach using
becomes less efficient. However, for all Xu -to-Xg ratios (1.65,
a synthetic example in accordance to the previously introduced
0.82, and 0.46), the extracted utility features [Fig. 5(d)–(f)]
field data. Xu and Xg represent the utility and geological back-
are again in good agreement with features an experienced
ground response, respectively, and Xm = Xu + Xg simulates a
interpreter would depict by visually inspecting the individual
migrated GPR time slice image. Calculation of phase symmetry
images [Fig. 5(a)–(c)].
S(Xm ) successfully enhances the target features from Xm as
well as symmetric features from Xg . In order to suppress the
C. Application to Field Data
geological background response, we use an image-processing
step called Laplacian-based highboosting [29] prior to the For the dual-configuration GPR data collected at the
calculation of the S-attribute. Thereby, we attenuate large-scale Horstwalde test site, we calculate S for each data set after
global features (introduced by Xg ) and enhance sharp local applying Laplacian-based highboosting. The information from
features. The result of applying Laplacian-based highboosting the X11 and X22 data is combined by
to Xm is shown in the panel labeled as L(Xm ). Subsequent cal-
culation of phase symmetry (S ◦ L)(Xm ) shows a significantly Sdc = max {(S ◦ L)(X11 ), (S ◦ L)(X22 )} (4)
improved result compared to S(Xm ). Feature continuity agrees to ensure maximum information from both components. The
well with the one obtained by visual inspection. application of a temporal running-average filter (window size
To further investigate the efficiency of phase symmetry for of 1.4 ns) has proven to further emphasize the expected tar-
utility extraction, we conducted a noise study using the syn- get features. Given the Sdc image, we calculate the feature
thetic example from Fig. 4. In doing so, we steadily decreased angle φf using Prewitt gradients [29] by
the amplitude ratio between Xu and Xg . Fig. 5 shows the   
 Sdc ∗ PE  π
results of this analysis for three selected Xu -to-Xg ratios. As φf = arctan − (5)
expected, by decreasing the Xu -to-Xg ratio, the resulting phase Sdc ∗ PN  4
BÖNIGER AND TRONICKE: SUBSURFACE UTILITY EXTRACTION AND CHARACTERIZATION 741

Fig. 5. Evaluation of the influence of the Xu -to-Xg ratio on the performance of phase-symmetry-based feature extraction. (a)–(c) GPR time slice images Xm
with decreasing Xu -to-Xg ratio from left to right. (d)–(f) Corresponding phase symmetry images obtained after Laplacian-based highboosting (S ◦ L)(Xm ).
The Xu -to-Xg ratios were calculated by dividing the root-mean-square amplitude of the utility signals by the standard deviation of the geological background.

Fig. 6. (a) Fused X11 and X22 data time slice at 21 ns, (b) corresponding phase symmetry Sdc , and (c) feature angle φf slices calculated using (4) and (5),
respectively.

with PN and PE being the convolutional Prewitt operators orientation of these objects. Later on, φf is used to calculate
sensitive along local northing and easting, respectively. Fig. 6 a new depolarization attribute, and Sdc is used to render the
shows the application of this procedure to a selected time slice target objects in 3-D.
at 21 ns (see also Fig. 1). For comparison, Fig. 6(a) shows
a migrated data slice after fusing the migrated data cubes
IV. PCA-D ERIVED P OLARIZATION ATTRIBUTES
X11 and X22 [Fig. 1(e) and (f)]. Fusion is performed using
a weighted summation, where the weights are based on the PCA is a common statistical tool in various scientific fields,
signal energy calculated over a window length corresponding including the analysis of geophysical data [31]–[35]. In the
to the dominant wavelength. This fusion approach minimizes following, we introduce the basics of PCA being relevant for
information loss (loss of sharpness) in a signal-driven manner. our analyses. For a detailed introduction into the field of PCA,
An alternative wavelet-based data fusion method can be found we would like to refer the interested reader to [36] and [37].
in [30]. It should be noted that the fused data are only used for PCA is based on the extraction of dominant features from
comparison, while all attribute analyses are performed using the the data covariance matrix. The eigenvalues of the covariance
individual X11 and X22 data. matrix, in our case from X11 and X22 , represent the so-
Comparing Fig. 6(a)–(c), we see that Sdc successfully called principal components. From these principal components,
highlights the target objects, i.e., three west–east and one several attributes can be derived. In this study, we focus on
north–south oriented utility pipes, while φf quantifies the the polarization angle φp and the linearity R. The polarization
742 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 50, NO. 3, MARCH 2012

Fig. 7. Four potential scenarios to illustrate the proposed attributes φp and R. The examples are calculated using Morlet wavelets to represent dual-configuration
GPR data (X11 , X22 ). (a) Identical (unpolarized) waveforms, (b) phase-shifted waveforms, (c) in-phase waveforms with different amplitudes, and (d) phase-
shifted waveforms with differing amplitudes.

angle [38] is the orientation of the two principal components


(pc1 and pc2 ) which we calculate by
  
 pc1  π

φp = arctan − . (6)
pc2  4

As used here, linearity (also termed as degree of rectilinearity)


has been introduced by Flinn [39]. In our case, it is defined as
 
min{λC,1 , λC,2 }
R=1− (7)
max{λC,1 , λC,2 }

with λC,i being the ith eigenvalue of the covariance matrix.

A. Synthetic Studies
To further illustrate φp and R to analyze GPR data, Fig. 7
shows the influence of amplitude and phase variations of a
Morlet wavelet on these attributes. In Fig. 7(a), the two compo-
nents/waveforms (X11 and X22 ) are identical and, thus, result
in φp = 0 (no polarization) and R = 1 (complete redundancy). Fig. 8. Influence of (a) waveform amplitude variations and (b) phase shifts on
In Fig. 7(b), we introduce a phase shift between the two the proposed attributes φp and R. The examples are calculated using Morlet
wavelets to represent dual-configuration GPR data (X11 and X22 ).
waveforms of π/2, which results in φp = 0.785 and R = 0.
Here, it is obvious that, for R close to zero (e.g., circular
B. Application to Field Data
polarization), φp is not well defined. Fig. 7(c) and (d) further
demonstrates the influence of amplitude and phase variations. For the dual-configuration GPR data collected at the
It can be seen that amplitude variations rotate φp while phase Horstwalde field site, we calculated the PCA-derived attri-
variations predominantly influence R. In order to investigate butes φp and R using (6) and (7) (Fig. 9). We calculated φp
the influence of amplitude and phase variations between two and R from the unmigrated [Fig. 9(a) and (b)] and migrated
waveforms, we calculated 2-D polarization and linearity maps [Fig. 9(c) and (d)] data cubes for a selected time slice at 21 ns
over varying amplitude and phase terms. Fig. 8(a) shows the (see Fig. 1) using a time window of 12 ns. Polarization angles
effects of amplitude variations between X11 and X22 and its [Fig. 9(a) and (c)] further quantify which data set dominates
influence on φp . φp clearly depicts the amplitude variations in specific areas, e.g., positive angles indicate that X11 am-
between X11 and X22 , while R is insensitive to such variations plitudes are significantly larger than the corresponding X22
[not shown in Fig. 8(a)]. Fig. 8(b) shows the influence of phase amplitudes. Linearity [Fig. 9(b) and (d)] enhances the target
variations between 0 and 2π on R. In contrast to amplitude features. The result is similar to an edge-enhancing attribute due
variations, phase variations do not influence φp [not shown to strong waveform changes at the pipe edges. Comparison of
in Fig. 8(b)] but have a significant influence on R, which is unmigrated and migrated attribute slices illustrates the focusing
indicated by the regular patterns in Fig. 8(b). These plots help to effect of migration, resulting in significantly sharper images.
understand the sensitivity of φp and R with respect to amplitude When comparing Figs. 7(b) and 9(c) and (d), it is evident that
and phase differences between two waveforms and represent the west–east oriented pipe at ∼12–13 m northing is not clearly
the basis to analyze our dual-configuration 3-D GPR data. imaged by R. For this object, negative φp values indicate
BÖNIGER AND TRONICKE: SUBSURFACE UTILITY EXTRACTION AND CHARACTERIZATION 743

Fig. 9. (a) φp and (b) R calculated over a time window length of 12 ns using the unmigrated data cubes. (c) φp and (b) R calculated over a time window length
of 12 ns using the migrated data cubes.

different polarization phenomena compared to the west–east the surrounding material, while the weak scattering of Pipe 2
oriented pipes at ∼10 and ∼14 m northing, respectively. Thus, is indicative for dielectric pipes with permittivities lower than
we conclude that the polarization attributes φp and R represent the surrounding material. Based on these observations, our data
efficient tools to highlight and summarize polarization effects suggest that at least two different types of pipes are present
usually present in multicomponent GPR data. at our site. Considering additional data (e.g., the results of an
EM induction survey; not shown here) and site information, we
V. D EPOLARIZATION A NALYSIS interpret Pipes 1, 3, and 4 as metal pipes, while Pipe 2 can
be interpreted as a nonmetallic high-impedance pipe, clearly
In order to combine the information from the two previously
differing from Pipes 1, 3, and 4 (see, e.g., [40]).
defined attributes (feature and polarization angle), we introduce
In order to validate this interpretation, we conducted excava-
a novel attribute, which we call depolarization Dp
 tions at four selected locations. Fig. 10(c) shows our excavation
φp /φf , if |φf | ≥ |φp | results, which confirm our GPR data interpretation. Pipes 1, 3,
Dp = (8)
φf /φp , if |φf | < |φp | and 4 (characterized by strong polarization effects) represent a
network of galvanized steel pipes with a diameter of ∼40 mm
where φf and φp are calculated using (5) and (6), respectively.
located at a depth range between 1.10 and 0.66 m. The weakly
Dp varies between −1 and 1. A Dp value of 1 indicates a
polarized event (Pipe 2) proved to be a sewer pipe made of
feature exhibiting maximum polarization with the long axes of
stoneware with a diameter of ∼180 mm buried at a depth of
the antenna dipoles parallel to the long axis of the feature. In
∼0.74 m below surface. The depths observed for Pipes 1 and
contrast, a Dp of −1 indicates a feature exhibiting maximum
2 differ from the extracted ones. Inspecting Figs. 3 and 10 in
polarization with the long axes of the antenna dipoles perpen-
detail, we see that the reflections from these pipes interfere
dicular to the long axis of the feature.
with the ground-wave arrivals. Detailed data analysis at the
Fig. 10 shows the final interpretation of our dual-
corresponding locations shows that this interference is predomi-
configuration data set, including extracted target depths
nantly shifting the peak energy of these pipe responses to larger
[Fig. 10(a)], calculated Dp attribute values [Fig. 10(b)], and
depths. The strong depth and positional match between GPR
available ground-truth information [Fig. 10(c)]. In Fig. 10(a)
and excavation results for Pipes 3 and 4 illustrate that using
and (b), rendering of all pipe features is based on iso-surfaces
higher antenna frequencies (resulting in a clearer separation
calculated from the 3-D phase symmetry cube [see also
between the ground-wave and the pipe reflections) would also
Fig. 7(b)]. Fig. 10(a) shows the extracted features with the top
improve the extracted depth locations of these pipes.
depth mapped onto the iso-surface. It can be seen that the pipes
are located at varying depths ranging from ∼1.11 to ∼0.95 m.
In Fig. 10(b), we mapped Dp onto the same iso-surfaces. This
figure illustrates how this attribute successfully combines the VI. D ISCUSSION AND C ONCLUSION
geometrical and physical information of the X11 and X22 data Using field data recorded across different utility pipes at a
sets. The pipes labeled as 1, 3, and 4 are characterized by strong well-constrained field site, we have demonstrated the potential
preferential scattering when the long axis of the dipole antennas of considering and analyzing polarization phenomena in 3-D
is oriented parallel to the orientation of the pipes, while Pipe 2 GPR data. We have introduced an attribute-based processing
is characterized by weak preferential scattering when the long flow to enhance the structural and physical characteristics of
axis of the antennas is orthogonal to the orientation of the pipe. the target objects. These attributes are calculated from dual-
Following the findings in [4], the observed strong polarization configuration 3-D GPR data consisting of two data volumes
effects of Pipes 1, 3, and, 4 are indicative for metal and low- recorded using parallel and perpendicular broadside antenna
impedance dielectric cylinders with permittivities higher than configurations, respectively.
744 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 50, NO. 3, MARCH 2012

Fig. 10. Final interpretation of the 3-D GPR field data. (a) Depth to the top pipe surface, calculated using a constant velocity of 0.111 m/ns, mapped onto
iso-surfaces extracted from the phase symmetry cube. (b) Depolarization attribute mapped onto the extracted iso-surfaces. (c) Ground-truth information obtained
from excavation. The white markers in (a) and (b) correspond to the pipe locations from the available ground-truth information.

Phase symmetry has proven to successfully delineate the objects. Combination of our structurally and physically based
objects of interest as this attribute highlights symmetrical fea- attributes into a single new attribute, called depolarization,
tures and also allows for volumetric feature extraction. Fur- allows one to distinguish between different types of subsurface
thermore, PCA has proven to be a feasible tool to extract the utility pipes based on their depolarization characteristics. This
dominant polarization characteristics from dual-configuration is a major improvement compared to standard GPR approaches
GPR data. We have calculated the PCA-based attribute po- largely relying on interpreting diffraction hyperbolas in single-
larization angle and linearity and found that these attributes component data sets. Thus, we have successfully developed an
are extremely useful to further characterize the target utility approach toward the inclusion of polarization characteristics of
BÖNIGER AND TRONICKE: SUBSURFACE UTILITY EXTRACTION AND CHARACTERIZATION 745

GPR wavefields into attribute-based analysis and interpretation [11] R. Streich and J. van der Kruk, “Accurate imaging of multicomponent
flows. GPR data based on exact radiation patterns,” IEEE Trans. Geosci. Remote
Sens., vol. 45, no. 1, pp. 93–103, Jan. 2007.
Given the recent availability of multicomponent sensor sys- [12] J. P. V. Gestel and P. L. Stoffa, “Application of Alford rotation to ground-
tems acquiring different antenna configurations simultaneously, penetrating radar data,” Geophysics, vol. 66, no. 6, pp. 1781–1792,
we believe that a combined interpretation, as developed in Nov./Dec. 2001.
[13] S. J. Seol, J.-H. Kim, Y. Song, and S.-H. Chung, “Finding the strike
this study, can result in significantly improved interpretation direction of fractures using GPR,” Geophys. Prospecting, vol. 49, no. 3,
for utility detection and characterization. In this study, we pp. 300–308, May 2001.
focused on the methodological introduction of the new analysis [14] P. D. Kovesi, “Invariant measures of image features from phase informa-
tion,” Ph.D. dissertation, Univ. Western Australia, Perth, Australia, 1996.
framework. At this stage, we have not included the influence of [15] U. Böniger and J. Tronicke, “Improving the interpretability of 3D GPR
the radiation characteristics of the source and receiver antenna data using target-specific attributes: Application to tomb detection,”
into our analysis. Streich et al. [41] recently demonstrated that, J. Archaeol. Sci., vol. 37, no. 4, pp. 672–679, 2010.
[16] U. Böniger and J. Tronicke, “On the potential of kinematic GPR survey-
for polarization-independent scatterers, the incorporation of ing using a self-tracking total station: Evaluating system cross-talk and
the radiation characteristics into a single- and multicomponent latency,” IEEE Trans. Geosci. Remote Sens., vol. 48, no. 10, pp. 3792–
3798, Oct. 2010.
vector-migration approach results in adequate GPR images. For [17] D. K. Butler, Near-Surface Geophysics: Investigations in Geophysics
polarization-dependent scatterers as presented in this study, the No. 13. Tulsa, OK: Soc. Exploration Geophysicists, 2005.
use of an exact-field vector-migration scheme is required to ef- [18] D. Watson, Contouring: A Guide to the Analysis and Display of Spatial
Data. New York: Pergamon, 1992.
ficiently address radiation-pattern and vector-wave-propagation [19] M. Sambridge, J. Braun, and H. McQueen, “Geophysical parametrization
effects. We thus believe that the application of an exact-field and interpolation of irregular data using natural neighbours,” Geophys. J.
vector-migration scheme would further improve the fidelity Int., vol. 122, no. 3, pp. 837–857, Dec. 1995.
[20] B. Heincke, W. Wheeler, J. S. Rnning, H. C. Nordheide, K. Blum, and
of our interpretation. Nevertheless, comparing the raw and E. Dalsegg, “Using 2-D and 3-D georadar to characterize karst collapse
migrated polarization attributes in our 3-D data set, no obvious as a basis for high-resolution reservoir-analogue models,” in Proc. 12th
variations between the raw and migrated data are observable Int. Conf. Ground Penetrating Radar, Birmingham, U.K., 2006.
[21] C. W. Tyler, Human symmetry perception and its computational analysis.
apart from the desired focusing effect. Thus, we believe that Mahwah, NJ: Lawrence Erlbaum Associates Publ., 2002.
our approach is capable in capturing the dominant polarization [22] J. Wagemans, “Detection of visual symmetries,” Spatial Vision, vol. 9,
effects. Application of this algorithm for automatic feature no. 1, pp. 9–32, 1995.
[23] T. S. Huang, J. W. Burnett, and A. G. Deczky, “The importance of phase in
extraction and classification might be a further focus of future image processing filters,” IEEE Trans. Acoust., Speech, Signal Process.,
work. Additionally, including cross-polarized or 45◦ broadside vol. ASSP-23, no. 6, pp. 529–542, Dec. 1975.
configurations into the surveying strategy might further im- [24] M. C. Morrone and R. A. Owens, “Feature detection from local energy,”
Pattern Recognit. Lett., vol. 6, no. 5, pp. 303–313, Dec. 1987.
prove the potential of our approach. Aside from utility detection [25] P. D. Kovesi, “Symmetry and asymmetry from local phase,” in Proc. 10th
and characterization, our strategy might also be applicable to Australian Joint Conf. Artif. Intell., 1997, pp. 185–190.
other GPR applications, e.g., the characterization of geological [26] P. D. Kovesi, “Image features from phase congruency,” Videre: A J. Com-
put. Vis. Res., vol. 1, no. 3, pp. 1–26, 1999.
targets such as fractures and faults. [27] D. J. Field, “Relations between the statistics of natural images and the
response properties of cortical cells,” J. Opt. Soc. Amer. A, Opt., Image
Sci., Vis., vol. 4, no. 12, pp. 2379–2394, Dec. 1987.
[28] B. Russell, D. Hampson, and J. Logel, “Applying the phase congru-
R EFERENCES ency algorithm to seismic data slices—A carbonate case study,” in Proc.
[1] D. J. Daniels, Ground Penetrating Radar, 2nd ed. London, U.K.: Peter Frontiers Innov.—CSPG CSEG CWLS Convention, 2009, pp. 677–680.
Peregrinus Ltd., 2004, ser. IEE Radar, Sonar, Navigation and Avionics [29] R. C. Gonzalez and R. E. Woods, Digital Image Processing, 2nd ed.
Series. Upper Saddle River, NJ: Prentice-Hall, 2002.
[2] H. M. Jol, Ground Penetrating Radar: Theory and Applications. [30] J. Hugenschmidt and A. Kalogeropoulos, “The inspection of retaining
Amsterdam, The Netherlands: Elsevier, 2009. walls using GPR,” J. Appl. Geophys., vol. 67, no. 4, pp. 335–344, Apr. 2009.
[3] R. L. Roberts and J. J. Daniels, “Analysis of GPR polarization phenom- [31] G. S. Wagner and T. J. Owens, “Signal detection using multi-channel
ena,” J. Environ. Eng. Geophys., vol. 1, no. 2, pp. 139–157, Aug. 1996. seismic data,” Bull. Seismol. Soc. Amer., vol. 86, no. 1A, pp. 221–231,
[4] S. J. Radzevicius and J. J. Daniels, “Ground penetrating radar polarization Feb. 1996.
and scattering from cylinders,” J. Appl. Geophys., vol. 45, no. 2, pp. 111– [32] J. R. Scheevel and K. Payrazyan, “Principal component analysis applied to
125, Sep. 2000. 3D Seismic data for reservoir property estimation,” SPE Reservoir Eval.
[5] G. P. Tsoflias, J. P. van Gestel, P. L. Stoffa, D. D. Blankenship, and Eng., vol. 4, no. 1, pp. 64–72, Feb. 2001.
M. Sen, “Vertical fracture detection by exploiting the polarization prop- [33] B. Karlsen, J. Larsen, H. B. D. Sorensen, and K. B. Jakobsen, “Compar-
erties of ground-penetrating radar signals,” Geophysics, vol. 69, no. 3, ison of PCA and ICA based clutter reduction in GPR systems for anti-
pp. 803–810, May/Jun. 2004. personal landmine detection,” in Proc. 11th IEEE Workshop Stat. Signal
[6] A. M. Thomas, C. D. F. Rogers, D. N. Chapman, N. Metje, and J. Castle, Process., 2001, pp. 146–149.
“Stakeholder needs for ground penetrating radar utility location,” J. Appl. [34] A. J. Gámez, C. S. Zhou, A. Timmermann, and J. Kurths, “Nonlinear
Geophys., vol. 67, no. 4, pp. 345–351, Apr. 2009. dimensionality reduction in climate data,” Nonlinear Process. Geophys.,
[7] L. Orlando and E. Slob, “Using multicomponent GPR to monitor cracks vol. 11, no. 3, pp. 393–398, 2004.
in a historical building,” J. Appl. Geophys., vol. 67, no. 4, pp. 327–334, [35] J. E. D. Pereira, A. J. Strieder, J. P. Amador, J. L. S. da Silva, and
Apr. 2009. L. L. V. Volcato, “A heuristic algorithm for pattern identification in large
[8] D. S. Sassen and M. E. Everett, “3D polarimetric GPR coherency multivariate analysis of geophysical data sets,” Comput. Geosci., vol. 36,
attributes and full-waveform inversion of transmission data for char- no. 1, pp. 83–90, Jan. 2010.
acterizing fractured rock,” Geophysics, vol. 74, no. 3, pp. J23–J34, [36] J.-L. Mari and D. Chapellier, Reservoir And Civil Engineering
May/Jun. 2009. Geophysics. Paris, France: Editions Technip, 1999.
[9] F. Lehmann, D. Boerner, K. Holliger, and A. Green, “Multicomponent [37] I. T. Jolliffe, Principal Component Analysis, 2nd ed. New York:
georadar data: Some important implications for data acquisition and pro- Springer-Verlag, 2002.
cessing,” Geophysics, vol. 65, no. 5, pp. 1542–1552, Sep./Oct. 2000. [38] P. N. Mahob and J. P. Castagna, “AVO hodograms and polarization
[10] J. van der Kruk, C. P. A. Wapenaar, J. T. Fokkema, and P. M. van den Berg, attributes,” Leading Edge, vol. 21, no. 1, pp. 18–27, Jan. 2002.
“Three-dimensional imaging of multicomponent ground-penetrating radar [39] E. A. Flinn, “Signal analysis using rectilinearity and direction of particle
data,” Geophysics, vol. 68, no. 4, pp. 1241–1254, Jul./Aug. 2003. motion,” Proc. IEEE, vol. 53, no. 12, pp. 1874–1876, Dec. 1965.
746 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 50, NO. 3, MARCH 2012

[40] J. L. Porsani, E. Slob, R. S. Lima, and D. N. Leite, “Comparing detection Jens Tronicke received the Diploma degree in geo-
and location performance of perpendicular and parallel broadside GPR physics from the University of Münster, Münster,
antenna orientations,” J. Appl. Geophys., vol. 70, no. 1, pp. 1–8, Jan. 2010. Germany, in 1997 and the Ph.D. degree from the Uni-
[41] R. Streich, J. van Der Kruk, and A. G. Green, “Vector-migration of stan- versity of Tübingen, Tübingen, Germany, in 2001.
dard copolarized 3D GPR data,” Geophysics, vol. 72, no. 5, pp. J65–J75, From 2001 to 2005, he was a Postdoctoral
Sep./Oct. 2007. Researcher with the Swiss Federal Institute of
Technology (ETH), Zurich, Switzerland. In 2002,
he received a visiting scientist fellowship by the
Urs Böniger (S’08–M’10) received the Diploma German science foundation at the Center of Geo-
degree in geophysics from the Swiss Federal Institute physical Investigations of the Shallow Subsurface
of Technology (ETH), Zurich, Switzerland, in 2004 at Boise State University. Since 2005, he has been
and the Ph.D. degree from the University of Potsdam, a Professor with the Institute of Geosciences, University of Potsdam,
Potsdam, Germany, in 2010. Potsdam, Germany. His research interests include georadar, electrical, and
He is currently with the University of Potsdam. seismic methods as well as the application of integrated geophysical approaches
His research focuses on attribute-based processing to environmental and hydrogeological site characterization.
and interpretation approaches for georadar data as Prof. Tronicke is a member of AGU, DGG, EAGE, EEGS, and SEG.
well as the application of seismic methods to char-
acterize a creeping landslide.

You might also like