Subsurface Classification Using Ground Penetrating Radar and Support Vector Machnines: A Survey

spraytownspeakerΤεχνίτη Νοημοσύνη και Ρομποτική

16 Οκτ 2013 (πριν από 3 χρόνια και 5 μήνες)

70 εμφανίσεις

UNIVERSITY OF PUNE







Subsurface Classification Using Ground

Penetrating Radar and Support Vector

Machnines: A Survey





by

Chetan Punekar






A thesis submitted in partial fulfillment for the

degree of M.Sc (Computer Science)



in the

Department of C
omputer Science & Engineering

Indian Institute of Technology Bombay






Acknowledgement
s


I would like to thank Dr. Umesh Bellur of Computer Science
& Engineering, and my other mentors and friends from the
Civil Engineering Department of the Indi
an Institute of
Technology Bombay, for their guidance and support
throughout this project
.












Index




Abstract



Subsurface characteristics and need for subsurface detection



Problems involved in
subsurface detection



Ground Penetrating Radar (GPR)



Supp
ort Vector Machines (SVM)



Motivation



Data Collection



Need for Computerization



Design Spec
ification (
Sequence
/ State Diagram
)



Limitations and Drawbacks

/ Challenges



Future Work



References & Bibliography










Abstract



The notion of subsurface detectio
n dates back from the late 70's.
Subsurface detection and classification can be carried out by
techniques like Seismic methods, Ground Penetrating Radar (GPR),
Nuclear techniques etc. The aim of our work is to extract and classify
meaningful subsurface cha
racteristics using the supervised Machine
Learning algorithm of Support Vector Machines (SVM). In our work
we make use of GPR to generate raw radargrams which are further
processed to remove clutter and noise using filtering techniques. The
extracted featu
res will be used to classify different soil types and
properties
.










Subsurface characteristics and need
for subsurface detection



The subsurface characteristics are the most

critical because design
engineers have little control over the conditions

he or she has to work
with beneath the ground surface.



There are many soil properties and groundwater parameters that
affect the classification of subsurface.

These include soil type, soil
permeability, grain size distribution, soil moisture content, p
H,

temperature, groundwater geochemistry, groundwater conductivity,
and depth to the groundwater.


Along with soil, there are many more things present underground,
for e.g. water pipes, voids, fossils, wires etc.

These materials

are part of the subsurface

and hence we cannot detect
them
. Any material has some properties which can be used to
distinguish it from other materials. These properties are



1)

Mechanical properties
-

compressive strength, ductility,
hardness, softness, tensile strength etc.

2)

Electri
cal properties


electrical conductivity, permittivity,
dielectric constant, dielectric strength etc.

3)

Thermal properties


thermal conductivity, thermal expansion,
specific heat, heat of vaporization etc.

4)

Chemical properties


pH, surface tension, corrosio
n resistance
etc.

5)

Magnetic properties


permeability, hysteresis etc.



The detection of buried or concealed objects is a subject of intense
interest in many areas of endeavor, including geological exploration
for minerals, characterization of sites used
for waste disposal, airport
security, mine and unexploded ordanance detection. Despite this
broad interest, robust detection of buried objects, classification of
subsurface, is still difficult to achieve in practice
.




















Problems involved i
n

subsurface
detection



Many approaches to subsurface detection have been formulated, but
all such approaches suffer from certain fundamental problems. For
example, many of the signals used to search for buried objects or to
classify the subsurface are p
artially reflected as they contact the air
-
earth interface. These signals may also be attenuated as they pass
through the earth. As a consequence, it can be difficult to propagate a
signal of adequate power to the object/material and back to the sensor.

In

addition to this, sensors receive spurious signals and noise from
various sources. Typical sources of noise include buried
inhomogeneities (rocks, voids, debris) and surface clutter (vegetation,
rocks, surface litter, topographic irregularities). In cert
ain cases the
detection system may confuse these undesired signals with those from
a target/material
.


For example, the metallic casing of a bomb is a much better
electrical conductor than rock or soil. The conductivity contrast
betw
een a buried item
and
its surroundings can be exploited by
several means in order to locate the item
. The difference in electrical
conductivity also means that an electromagnetic wave will be
reflected differently from a buried metallic item than from, say, a
rock
.


Ground Pene
trating
Radar (
GPR)



Ground penetrating radar (
GPR
) is a widely used
non
-
destructive
tool for the

investigation of the shallow subsurface, and is

particularly
useful in the detection and mapping

of subsurface utilities and other
solid objects
.



GPR has

been developed over the past thirty years for shallow, high
resolution investigations of the subsurface. GPR is a time
-
dependent
geophysical technique that can provide a 3
-
D pseudo image of the
subsurface, including the fourth dimension of color, and can
also
provide accurate depth estimates for many common subsurface
objects. Under favorable conditions, GPR can provide precise
information concerning the nature of buried objects. It has also proven
to be a tool that can be operated in boreholes to extend t
he range of
investigations away from the boundary of the hole.



GPR uses the principle of scattering of electromagnetic waves to
locate buried objects. The basic principles and theory of operation for
GPR have evolved through the disciplines of electrica
l engineering
and seismic exploration, and practitioners of GPR tend to have
backgrounds either in geophysical exploration or electrical
engineering. The fundamental principle of operation is the same as
that used to detect aircraft overhead, but with GPR
that antennas are
moved over the surface rather than rotating about a fixed point. This
has led to the application of field operational principles that are
analogous to the sei
smic reflection method



Working of
GPR
:



The practical result of the radiatio
n of electromagnetic waves into
the subsurface for GPR measurements is shown by the basic operating
principle that is illustrated in Figure A1. The electromagnetic wave is
radiated from a transmitting antenna, travels through the material at a
velocity whi
ch is determined primarily by the permittivity of the
material. The wave spreads out and travels downward until it hits an
object that has different electrical properties from the surrounding
medium, is scattered from the object, and is detected by a recei
ving
antenna. The surface surrounding the advancing wave is called a
wave
-
front
. A straight line drawn from the transmitter to the edge of
the wavefront is called a
ray
. Rays are used to show the direction of
travel of the wavefront in any direction away f
rom the transmitting
antenna. If the wave hits a buried object, then part of the waves
energy is “reflected” back to the surface, while part of its energy
continues to travel downward. The wave that is reflected back to the
surface is captured by a receive

antenna, and recorded on a digital
storage device for later interpretation.


Antennas can be considered to be transducers that convert electric
currents on the metallic antenna elements to transmit electromagnetic
waves that propagate into a material. A
ntennas radiate
electromagnetic energy when there is a change in the acceleration of
the current on the antenna. The acceleration that causes radiation may
be either linear,(e.g., a time
-
varying electromagnetic wave traveling
on the antenna), or angular ac
celeration. Radiation occurs along a
curved path, and radiation occurs anytime that the current changes
direction (e.g. at the end of the antenna element). Controlling and
directing the radiation from an antenna is the purpose of antenna
design.


Antenna
s also convert electromagnetic waves to currents on an
antenna element, acting as a receiver of the electromagnetic radiation
by capturing part of the electromagnetic wave. The
principle of
reciprocity
says that the transmit and receive antennas are
interc
hangeable, and this theory is valid for antennas that are
transmitting and receiving signals in the air, well above the surface of
the ground. In practice, transmit and receive antennas are not strictly
interchangeable when placed on the ground, or a lossy

material
surface, because of attenuation effects of the ground in the vicinity of
the transmit antenna.


Electromagnetic waves travel at a specific velocity that is
determined primarily by the permittivity of the material. The
relationship between the v
elocity of the wave and material properties
is the fundamental basis for using GPR to investigate the subsurface.
To state this fundamental physical principle in a different way: the
velocity is different between materials with different electrical
propert
ies, and a signal passed through two materials with different
electrical properties over the same distance will arrive at different
times. The interval of time that it takes for the wave to travel from the
transmit antenna to the receive antenna is simply
called the
travel
time
. The basic unit of electromagnetic wave travel time is the
nanosecond (ns), where 1 ns = 10
-
9
s. Since the velocity of an
electromagnetic wave in air is 3x10
8
m/s (0.3 m/ns), then the travel
time for an electromagnetic wave in air is

approximately 3.3333 ns
per m traveled. The velocity is proportional to the inverse square root
of the permittivity of the material, and since the permittivity of earth
materials is always greater than the permittivity of the air, the travel
time of a wav
e in a material other than air is always greater than
3.3333 ns/m. The travel time of an electromagnetic wave through two
different materials is shown in

Figure
.




Considering the wave scattered from the object in Figure A1, if a
receive antenna is swit
ched
-
on at precisely the instant that the pulse is
transmitted, then two pulses will be recorded by the receive antenna.
The first pulse will be the wave that travels directly through the air
(since the velocity of air is greater than any other material),
and the
second pulse that is recorded will be the pulse that travels through the
material and is scattered back to the surface, traveling at a velocity
that is determined by the permittivity (e) of the material. The resulting
record that is measured at the

receive antenna is similar to one of the
t
ime
-
amplitude plots in Figure
(b), with the “input” wave consisting
of the direct wave that travels through air, and the “output” pulse
consisting of the wave reflected from the buried scattering body. The
recordi
ng of both pulses over a period of time with receive antenna
system is called a “trace”, which can be thought of as a time
-
history
of the travel of a single pulse from the transmit antenna to the receive
antenna, and includes all of its different travel pa
ths. The
trace
is the
basic measurement for all time
-
domain GPR surveys. A
scan
is a
trace where a color scale has been applied to the amplitude values.
The round
-
trip (or two
-
way) travel time is greater for deep objects
than for shallow objects. Therefore
, the time of arrival for the
reflected wave recorded on each trace can be used to determine the
depth of the buried object, if the velocity of the wave in the
subsurface is

known.


GPR energy responds to different materials in different ways. It is
gover
ned by two physical properties of the material


electrical
conductivity, dielectric constant.

The strength of a reflection is
proportional to the dielectric constant between the two materials. The
greater the contrast, the brighter the reflection.








Section of data with layer (concrete bottom) and target (rebar)
reflecti
on (
.dzt file)



Boundary

Dielectric contrast

Reflection stre
ngth

Asphalt
-
concrete

Medium

Medium

Concrete
-
sand

Low

Weak

Concrete
-
air

High, phase reversal

Strong

Concrete deck
-
concrete beam

None

No Reflection

Concrete
-
metal

High

Strong

Concrete


water

High

Strong

Concrete


PVC

Low to Medium,
phase reversal

Weak







Sample
Radar grams
:












S
upport Vector
Machine (
SVM)




SVM is a
supervised learning technique from the field of Machine
Learning
. It is a
pplicable to both classification and regression

and

is

b
ased on the principles of Structural Ris
k Minimization
.


SVMs are used for Classification, Regression, Supervised and
Unsupervised Learning
.

There are two key concepts of a SVM:



A
maximal margin classifier

-

It is a linear classifier
which
constructs a separating hyper plane to maximize distanc
e
between data.



A kernel function

to map the data in a new space where they are
separable (i.e. SVM is a kernel method).


Example: Linear SVM


Kernel Function:




Maps the data into high dimensional space in order to


detect the structure of data more easily



It returns the inner product between the data points



Choi
ce of
the kernel depends on the
application



Examples of Kernels:



Linear:
K
(
x
i
,
x
j
)=
x
i
T
x
j




Polynomial of power
p
:
K
(
x
i
,
x
j
)= (1+
x
i
T
x
j
)
p



Gaussian (radial
-
basis function network:








Sigmoid:
K
(
x
i
,
x
j
)= tanh(
β
0
x
i
T
x
j
+
β
1
)





Selection of Kernel:



The selection of the kernel depends upon the data type and accuracy
of the correctly classified objects.


The higher the accuracy rate, the better the kernel is for
classification.



Letter

example
:
-

Number

of classes: 26

Numbe
r

of data: 15,500 / 5000
(testing) / 10,500 (training)

Number

of features: 16




Kernel

Used
: L2
-
regularized L2
-
loss support vector
classification (dual)



Cost parameter: 50



Accuracy = 66.0200%


(3301/5000)




Kernel: multi
-
class support vector classification
by Crammer
and Singer



Cost parameter: 50



Accuracy = 76.8800%

(3844/5000)




Kernel:L1
-
regularized L2
-
loss support vector classification



Cost parameter: 50



Accuracy = 66.1600%


(3308/5000)




Kernel: polynomial: (gamma*u'*v + coef0)^degree



Cost parameter: 50



Accuracy = 81.24%


(4062/5000)


(classification)




Kernel: radial basis function: exp(
-
gamma*|u
-
v|^2)



Cost parameter: 50



Accuracy =
91.28%


(4564/5000)


(classification)




Kernel: sigmoid: tanh(gamma*u'*v + coef0)



Cost parameter: 50



Accuracy = 77.52%


(387
6/5000)

(classification)






Motivation




The development of an integrated soil survey information system
and needs for technical and scientific soil services require more
detailed and site
-
specific information concerning the properties,
composition, a
nd variability of soils. Frequently, information is
required from zones deeper than the limits of modem pedologic
investigations or to depths where insufficient observations have been
made to establish reliable standards. To fulfill these needs, different
methods of observing subsurface soils are

required.















Data Collection



A
GPR survey

was conducted

ne
ar Central Library, IIT Bombay.

The raw data gathered had many impurities like noise, air

etc.

All the
impurities were removed using the RADA
N

software and data was
made available for further processing.


The data obtained

was available in .dzt file format(RADAN file
format) and hence had to be converted into

appropriate file format for
classification.

Hence, we converted the .dzt file into AS
CII file
.

Also, the .dzt file can be converted into .
bmp (
bitmap) image
.








.
bmp image of
bore well

data



The known subsurface of the
bore well

and the
corresponding el
ectromagnetic pulse






Problem
Definition




The classification of subsurface is a challenging task.

The GPR
transmits electromagnetic pulses and receives back these
electromagnetic pulses back based on the dielectric contrasts between
the materials. By

looking at the .dzt file

(radargram)

we can identify
some objects by their pattern, for e.g. If we see hyperbolas, then it
indicates the presence of pipes or some circular objects
, but other than
this,
we don’t have enough information about the sub surfac
e.

The GPR is mostly used for detection of objects, materials etc.

What we have tried here is to classify the sub surface using machine
learning algorithms.

Classification of sub sur
face would help to
know the sub surface prior
before

the excavation.












Need for Computerization




The sub surface is unknown to
us;

the only means to learn about the
sub surface is excavation. However, the GPR electromagnetic pulse
s
can penetrate the sub surface and the GPR control unit stores the data.
These data fi
les are huge in numbers and it
requires highly skilled and
experienced professionals

to classify the

data by just observing the
radargram
,

this induces a lot of cost

and
the o
nly thing we learn from
the radargram

is: change in the soil layer, detection of
some object
etc.


Hence,
to reduce the cost of the project and to classify bulk data
within short time and also to increase the accuracy of the
classification,

there is need for computerization.
















Design Specification




Sequence Diagram





































Limitations and Drawbacks/
Challenges




The GPR responds only to the changes in the dielectric constants and
hence the radargram is an image based only on these dielectric
contrasts. Als
o the dielectric constants of som
e materials f
all with
in
some range
, for e.g. cement

(plain)


1.5
-
2.1, cement

(powder)


5
-
10
,
charcoal


1.2


1.81. Now the dielectric of charcoal is overlapped by
the dielectric of cement, so if we derive the dielectric of a material,
there can be a pos
sibility of wrongly classifying the material, like in
our case we can predict that the material is charcoal but in reality it
would be cement.


Radar energy responds to different materials in different ways. The
way that it responds to each

material is g
overned by two physical
properties of the material. The first one is
electrical

conductivity

and
the second one is
dielectric constant
, hence initially we have to derive
the dielectric constants of the materials and as the dielectric constants
are ambiguou
s we have to derive some more properties of the material
based on the dielectric constants
.


There are many soil properties and groundwater parameters that
affect the classification of subsurface. These include soil type, soil
permeability, grain size di
stribution, soil moisture content, pH,
temperature, groundwater geochemistry, groundwater conductivity,
and depth to the groundwater
, hence the classification of two same
subsurfaces present at two different locations would differ, i.e. if both
the surface
s are made of same material, they would be classified
differently based on the above parameters. This may also be true for
the same subsurface in different seasons.
























Future Work




The classification of subsurface using Ground Penet
rating Radar is a
very tedious and complicated procedure. The complexity is increased
because the GPR works on the principle of
dielectric contrasts
. To
identify the patterns in the radargram by naked eye, there is a need of
highly skilled and experienced
professionals, which incurs lot of cost.


Currently, we have tried to classify the subsurface soil types based
on the dielectric contrasts. The bigger picture can be classification of
each material, object, composition of the subsurface etc.















References and Bibliography



The

A
utomatic

D
etermination

of

S
oil

P
ermittivity

U
sing

T
he

R
esponse

F
rom

a

S
ubsurface

L
ocal

O
bject.
-

Golovko M. M.


Toward an Op
timal SVM Classification System
for

Hyperspectral Remote Sensing Images
.
-

Yakoub Bazi


Use of
soil information to determine application of ground

penetrating radar
.
-

James A. Doolittle a, Mary E. Collins


Automatic detection of buried utilities and solid objects with
GPR

using neural networks and pattern recognition
.
-


W. Al
-
Nuaimy, Y. Huang , M.

Nakhkash , M.T.C. Fang, V.T.
Nguyen,

A. Eriksen


Use of Dielectric Constant Reflection

Coefficients for
Determination of Groundwater Using Ground Penetrating
Radar. A.D.Vasudeo.
-
Y.B.Katpatal and R.N.Ingle


Application of Fuzzy
-
Neural Network in Classifi
cation of

Soils using G
round
-
penetrating Radar Imagery.
-

Lameck O.
Odhiambo
,
D. Tech. Sc.
Robert S. Freeland
,
Ph.D.. P.E
.
Post
-
doctoral Research Associate Associate Professor

Ronald E. Yoder
,
Ph.D., P.E.
J. Wesley Hines
,
Ph.D., MBA

Professor and Head,
Associate Professor


Soil Investigations using Electromagnetic Inducti
on and
Ground
-
Penetrating Radar
in Southwest Tennessee

Daniel

J. Inman, Robert S. Freeland,
John T. Ammons, and
Ronald E. Yoder


Application of 3D visualization

techniques in the analysi
s

of
GPR data for archaeology

Luigia Nuzzo,
-

Giovanni Leucci,
Sergio Negri, Maria Teresa Carrozzo and Tatiana Quarta

GPR based soil electromagnetic parameters determination for

subsurface imaging
.
-

R. Solimene, G. Prisco, and F. Soldovieri


Seismic refle
ction and ground
-
penetrating radar

imaging of a
shallow aquifer.
-

Steven J. Cardimona, William P. Clement
z
,

and Katharine Kadinsky
-
Cade


Classification System for Ground Penetrating Radar

Parameters
.
-

Richard Yelf, Daniel Yelf and Waleed Al
-
Nuaimy


Effe
cts of soil physical properties on GPR for

landmine
detection
.
-

Timothy W. Miller, Brian Borchers, Jan M.H.
Hendrickx, Sung
-
Ho Hong, Louis W. Dekker, and

Coen J. Ritsema


Modeling Dielectric
-
constant values of Geologic Materials: An
Aid

to Ground
Penetrat
ing Radar Data Collection and
Interpretation
.
-

Alex Martinez and Alan P. Byrnes


Applications of Ground Penetrating Radar (GPR) in

Detection
of Groundwater Table
.
-

Sabbar Abdallah Salih


Soil moisture content estimation using ground
-
penetrating

radar
ref
lection data
.
-

I.A. Lunt
a,
S.S. Hubbard
b
, Y. Rubin


Methods for prediction of soil dielectric properties: a review
.
-


Remke L. van Dama, Brian Borchersb, Jan M.H. Hendrickxa


Measuring the Electrical Properties of Soil Using a

Calibrated
Ground
-
Coupled G
PR System
.
-

C. P. Oden,
G. R. Olhoeft, D.
L. Wright, and M. H. Powers


Ground Penetrating Radar Fundamentals.
-

Jeffrey J. Daniels
.


GSSI Handbook For

RADAR Inspection of Concrete
.
-

Geophysical Survey Systems, Inc.


Automatic Detection and Classification

of Buried Objects

in GPR Images using Genetic Algorithms and Support Vector
Machines
.
-

Edoardo Pasolli, Farid Melgani, Massimo Donelli
,

Redha Attoui, and Mariette De Vos.


A Tutorial on Support Vector Machines for Pattern

Recognition
.
-

CHRISTOPHER J.C.
BURGES


Landmine Feature Extraction and Classification of

GPR Data
Based on SVM Method
.
-

Jing Zhang
, Liu Qun
, and Baikunth
Nath
.


LIBLINEAR: A
Library for Large Linear Classifi
cation
.
-


Rong
-
En Fan
,
Kai
-
Wei Chang
,
Cho
-
Jui Hsieh
,
Chih
-
Jen Lin
.


LIBSVM: a
Library for Support Vector Machines
.


Ch
ih Chung
Chang and Chih Jen Lin.


Support Vector

Machines for

Pattern

Classification
.


Shigeo
Abe
.