Vous êtes sur la page 1sur 4



Edoardo Pasolli (1), Farid Melgani (1), Massimo Donelli (1),
Redha Attoui (2), and Mariette De Vos (2)

Dept. of Information Engineering and Computer Science, Univ. of Trento,

Via Sommarive, 14, I-38100 Trento, Italy
E-mail: melgani@disi.unitn.it
Dept. of Philosophy, History and Cultural Heritage, Univ. of Trento,
Via Santa Croce, 65, I-38100 Trento, Italy

This work presents a novel pattern recognition approach for
the automatic analysis of ground penetrating radar (GPR)
images. The developed system comprises pre-processing,
segmentation, object detection and material recognition
stages. Object detection is done using an innovative
unsupervised strategy based on genetic algorithms (GA) that
allows to localize linear/hyperbolic patterns in GPR images.
Object material recognition is approached as a classification
issue, which is solved by means of a support vector machine
(SVM) classifier. Results on synthetic images show that the
proposed system exhibits promising performances both in
terms of object detection and material recognition.
Index Terms - Ground penetrating radar, buried
objects, genetic algorithms, support vector machine.
The development of noninvasive techniques to extract
information about the underground has shown in the last
years a growing interest by public as well as private entities
related to different application fields, such as oil and gas
exploration, geology, conduits and pipes location, and
archaeology. Depending on the application, an appropriate
sensor is used for imaging the underground. In particular, for
the problem of detecting buried objects at small depths,
which is the focus of this work, the most frequently used
technique is based on the ground penetrating radar (GPR).
This technique consists in the transmission and reception of
electromagnetic waves by means of which it is aimed at
achieving an exploration depth of few meters with a
resolution of several centimeters. Typically, the
interpretation of the large amount of acquired and stored
GPR data requires a human operator with high skill and
experience, involving thus high costs in terms of time and

978-1-4244-2808-3/08/$25.00 2008 IEEE

money. As a consequence, these cost problems have

encouraged an increasingly growing demand for the
development of automated subsurface mapping techniques
that are both accurate and rapid.
In the literature, there are still few published works
dealing with the automatic detection of patterns associated
with buried objects. In [1], the classical Hough transform
was used in order to identify linear segments in the image,
representing transitions between layers of different electrical
impedances. Authors proposed also a method for extracting
hyperbolic signatures of buried objects and hence estimating
their position. In [2], the detection process was subdivided
in three main stages: 1) preprocessing step to reduce noise
and undesired system effects; 2) image segmentation with an
artificial neural network classifier to identify areas
potentially containing object reflections; and 3) Hough
transform to detect hyperbolic patterns. In [3], some
preprocessing steps aiming at enhancing the signature of
buried targets were implemented. Then, automatic image
interpretation was carried out by a detector based on
artificial neural networks. In [4], authors applied a fuzzy
clustering approach to identify hyperbolas from GPR images
beforehand de-noised.
The automatic analysis of buried objects implies the
extraction of four main characteristics associated with these
objects, which raise the following four problems,
respectively: 1) object detection and localization; 2) object
material recognition; 3) object dimension estimation; and 4)
object shape recognition.
The objective of this paper is to propose a novel system
for the automatic analysis of GPR images capable of
identifying and classifying buried objects, i.e. for coping
with the first two above-mentioned problems. The developed
methodology comprises pre-processing, segmentation,
object detection and material recognition stages. Results on
synthetic images show that the proposed system exhibits
promising performances both in terms of object detection

II - 525


and material recognition.

characteristics of the ground (in particular the ground

permittivity value) and the acquisition parameters (i.e.,
sensor position and distance between transmitting and
receiving antennas).

2.1. Pre-processing
A pre-processing procedure is implemented for solving three
main issues: 1) reducing noise; 2) eliminating the undesired
presence of the ground surface echo; and 3) compensating
propagation losses. Noise reduction is performed with a
median filter, while the elimination of the ground surface
echo is done by a simple average operation. A time-gain
filter is used to compensate signal amplitude, attenuated by
spreading and losses.
2.2. Segmentation
The resulting pre-processed image is subject to a modulus
operator for overcoming phase inversion problems, which
are generated by the presence of objects with particular
dielectric characteristics. Then the image is thresholded to
discriminate between objects and the background. This
binarization operation, which allows to put under light the
parts of the image containing potential targets, is based on
the fact that buried objects are generally associated with
relatively large amplitude echoes. It is implemented by
means of the Kapurs thresholding technique [5]-[6], which
relies on the entropy maximization principle.
2.3. Single object detection
The next step consists in identifying the targets in the
obtained binary image in a completely unsupervised way.
This is done by means of a search of linear and hyperbolic
patterns representing potential targets. This search problem
is viewed as a matching problem in which it is looked for the
set of best linear or/and hyperbolic patterns fitting the
content of the binary image. It is solved through a genetic
optimization framework where the chromosome models
apex position and curvature coefficient associated with the
candidate pattern. Each gene assumes a real value spanning
an interval depending on image dimension. The adopted
fitness function is the Hamming distance between the
content of the binary image and the image that contains the
pattern encoded by the candidate chromosome.
2.4. Multiple object detection
Since the image may contain several patterns, the GA is run
in cascade several times, each associated with the search of a
single pattern in the image. Once a pattern is identified, it is
removed from the image to allow the GA for searching for
another pattern in the next iteration. The process is stopped
when a new extracted pattern is statistically incompatible
with the previous ones. The last phase consists in converting
detected patterns into geospatial coordinates, i.e., in
localizing the object in the ground in terms of horizontal
position and depth from the ground surface. The conversion
from image to geospatial domain is possible knowing the

After the object detection, the next phase consists in
recognizing the material of localized objects. We solve this
problem by viewing it as a classification issue which will be
based on the analysis of the waveform of the received
signals and will require the definition of two main
components, namely a feature extraction strategy and a
Since the apex position estimation performed by the
previously described detection method may be subject to
errors, a correction operation based on pattern energy
matching is first carried out. The feature extraction starts
hence from the knowledge of the apex position of the
detected pattern. In particular, a windowing operation
centered at the apex is applied for extracting the most
relevant signal amplitudes of the object. Afterwards, noise
effect is reduced through a mean operation performed over
adjacent traces. The result is a vector of waveform features,
each corresponding to a precise sampling time around the
object apex. Finally, the feature vector is normalized in the
range [0,1]. Thanks to normalization, the object features do
not depend neither on the dimension nor on the depth of the
object, but only on its dielectric characteristics.
Classification operation is performed by means of a
support vector machine (SVM) classifier [7]-[8]. The choice
of SVMs is motivated by their good generalization
capability and their low sensitivity to the curse of
dimensionality supported by the margin maximization
principle they are based on and their sparse representation of
the decision function.
In order to validate the proposed system, several sets of
experiments were conducted according to three phases
devoted to the assessment of: 1) the linear/hyperbolic pattern
detection method; 2) the material recognition method; and 3)
the global system, respectively.
In the first phase, 33 synthetic binary images conveying
hyperbolic and linear patterns in presence of noise were
produced. For each image, the parameters associated with
the patterns (i.e., number of patterns, position, curvature,
length, and thickness) were chosen randomly. Binary noise
was generated with a variable.
For the others phases, GPR images were generated
using the software GprMax [9]. This tool, developed using
the Finite Difference Time Domain (FDTD) method, allows
to simulate both two and three-dimensional acquisition
scenarios. In our case, two-dimensional scenarios were

II - 526

simulated using the GprMax2D version. Simulations were

designed by focusing on the archaeological application. We
fixed the signal frequency at 400 [MHz], the transmitterreceiver distance at 0.66O0 (being O0 the wavelength of the
electromagnetic wave) and the GPR position at 1.32O0 (i.e, 1
[m]) from the ground. Ground was assumed to be of sandy
nature. Therefore, its dielectric permittivity was equal to 4
and its conductivity was fixed to 0.01 [S/m]. While in the
second phase, the ground was supposed perfectly
homogeneous, in the third one we introduced in the
propagation medium clutter through scatterers, with random
spatial and intensity distribution. In the second phase, in
order to consider different acquisition scenarios, we
generated different GPR images by varying the number of
buried objects, their position, their size, their shape and their
material type. In particular, three types of shape (circular
section, square section, and uniform layer) and three types of
material (limestone, metal, and air) were considered. A
random Gaussian noise with variable intensity was also
added to the images. In the third phase, GPR images related
to 10 particular archaeological scenarios were also generated
in order to simulate acquisition scenarios that are as most
realistic as possible.
5.1. Detection method
Performances of the detection method were evaluated in
terms of horizontal and vertical pattern position errors. On
an average of all the test images, the horizontal apex
position has an error equal to 0.25 pixels, while the vertical
one is affected by an error equal to 0.95 pixels. The
asymptote slope, representing the hyperbola curvature, has a
mean error equal to 0.95. The developed multiple object
detection strategy allowed to detect correctly all the 165
patterns available in the images without any false pattern
detection despite the presence of noise.
5.2. Material recognition method
The classification accuracy of the material recognition
method was assessed by varying the type of kernel function
(linear or RBF) adopted for the SVM classifier, the number
of features (21, 41 or 61), and the number of training
patterns (300, 600 or 900). The best accuracy was obtained
using the RBF kernel function, a number of features equal to
61 and a number of training patterns equal to 900.
Moreover, some experiments to test the sensitivity of the
system to multiple object interference were also carried out.
As expected, best performances were obtained when just a
single object is present in the ground, but the system

exhibited a very satisfactory accuracy even when objects are

very close to each other (t0.125O0, i.e., around 9 [cm]).
5.3. Global system
Performances of the whole system were evaluated by
considering GPR images associated with 40 random
scenarios. In 25 scenarios (62%), all the buried objects
present in the corresponding image were correctly detected.
Simultaneous correct detection and material recognition of
all the objects was possible in 18 scenarios (45%). The 40
scenarios were characterized by a total number of buried
objects equal to 120. Considering separately the 120 buried
objects, in 85 cases (71 %) the objects were correctly
detected while simultaneous detection and material
recognition were obtained in 70 cases (58%). For the 85
correctly detected objects, obtained performances are as
follows: i) the mean error of the horizontal position is equal
to 6.42 [cm]; ii) the one of the vertical position is equal to
2.28 [cm]; and iii) the overall classification accuracy is
equal to 82%.
5.4. Application to archaeology
Finally, the system was applied to 10 archeological
scenarios. In general, the proposed method allowed to deal
well with them. An example is illustrated in Figure 1. In
particular, Fig.1-a shows the ground-truth which is
composed of a limestone block, that represents a staircase
made of three distinct steps. The related GPR image
generated by the GprMax software is represented in Fig.1-b,
where we can observe the hyperbolic patterns associated
with buried objects and contaminated by the presence of
noise. The result of the pre-processing stage is shown in
Fig.1-c, while the segmentation result is illustrated in Fig.1d. We notice that the three hyperpolic patterns associated to
the staircase (one for each step) are either completely or
partially captured by the binarization operation. The
detection method allowed to find correctly two of the three
patterns while a slight position shift was incurred for the
third one Fig.1-e. The material recognition stage associated
correctly the limestone class with all three patterns Fig.1-f.
At last, Fig.1-g depicts the reconstructed scenario which
shows that the three steps of the limestone block were
identified with a good positioning. In particular the depth of
the peak of the staircase was localized with a very good
precision. Just for one step, the horizontal position
determined by the system differs from the actual position of
some tens of centimeters. Their material type (limestone)
was correctly recognized.

II - 527








Fig. 1: Example of system evaluation on an archaeological scenario. (a) Simulated scenario; (b) Original GPR image; (c) Pre-processed
image; (d) Segmented image; (e) Segmented image with detected patterns; (f) Segmented image with detected patterns and their material
type; (g) Reconstructed scenario.

In this work, an innovative system for the automatic analysis
of GPR images has been presented. This system allows to
detect the presence in the ground of buried objects and to
estimate their position as well as their material type. After
the pre-processing and segmentation stages, the detection
operation is performed using a new iterative process based
on genetic algorithms, while the estimation of the material
type is handled as a classification issue solved by means of a
support vector machine classifier.
The obtained experimental results show that the
performances, in terms of both detection/positioning (error
of the order of few centimeters) and material recognition
(accuracy around 80%). The system appears efficient for the
automatic interpretation of GPR images in archaeological
Work is in progress to improve further the implemented
system by considering additional characteristics that we have
not examined in this work, namely the size estimation and
the shape recognition of the detected buried objects.
[1] L. Capineri, P. Grande, J. A. G. Temple, Advanced ImageProcessing Technique for Real-Time Interpretation of GroundPenetrating Radar Images, International Journal of Imaging
Systems and Technology, vol. 9, pp. 51-59, 1998.

[2] W. Al-Nuaimy, Y. Huang, M. Nakhkash, M.T.C. Fang, V.T.

Nguyen, A. Eriksen, Automatic Detection of Buried Utilities
and Solid Objects with GPR using Neural Networks and
Pattern Recognition, Journal of Applied Geophysics, vol. 43,
pp. 157-165, 2000.
[3] P. Gamba, S. Lossani, Neural Detection of Pipe Signatures in
Ground Penetrating Radar Images, IEEE Trans. Geosci. and
Remote Sens., vol. 38, pp. 790-797, 2000.
[4] S. Delb, P. Gamba, D. Roccato, A Fuzzy Shell Clustering
Approach to Recognize Hyperbolic Signatures in Subsurface
Radar Images, IEEE Trans. Geosci. and Remote Sens., vol.
38, pp. 1447-1451, 2000.
[5] J. N. Kapur, P. K. Sahoo, A. K. C. Wong, A New Method for
Graylevel Picture Thresholding using the Entropy of the
Histogram, Comput. Vis. Graph. Image Process., vol. 29, pp.
273285, 1985.
[6] F. Melgani, Y. Bazi, Markovian Fusion Approach to Robust
Unsupervised Change Detection in Remotely Sensed
Imagery, IEEE Geosci. and Remote Sens. Lett., vol. 3, pp.
457-461, 2006.
[7] V. Vapnik, Statistical Learning Theory. New York: Wiley,
[8] Y. Bazi, F. Melgani, Toward an Optimal SVM Classification
System for Hyperspectral Remote Sensing Images, IEEE
Trans. Geosci. and Remote Sens., vol. 44, pp. 3374-3385,
[9] A. Giannopoulos, GprMax: a Ground Penetrating Radar
Simulation Tool, http://www.gprmax.org, 2006.

II - 528