I. Introduction
An UXO classification technology based on fully polarimetric ground penetrating radar signatures has been developed since 1998. This surface-based radar system is operated in the frequency range from 20 MHz to 800 MHz. Radar data were collected near each “hot spot” selected from magnetic sensors that are effective in detecting subsurface metallic objects up to one meter depth. Multiple radar signatures were obtained from each metallic anomaly to determine whether it is an UXO or a false alarm. These radar signatures include electromagnetic resonances [1] and polarization properties, which can be directly or indirectly related to the geometry features such as length, length-to-width relationship, depth and orientation. Such a technology was demonstrated in January 2000 at the Tyndall AFB UXO test site and the results were reported in the previous forum [2]. The results obtained from this sandy site not only provided a baseline classification performance but also significantly improved the understanding of technology limitations as far as successful classification rate and false alarm rate are concerned [3]. These findings have led to important improvements in system configurations, measurement approaches and processing techniques. This paper discusses these improvements using measurement results.
II.
UXO Classification Based on Radar Signatures
The natural resonance feature associated with a conducting target is a result of the reverberation of surface currents induced by the incoming radar signal. Each reverberation is accompanied by radiations that damp the resonance strength. Further damping is caused by propagation through conducting soils. When these radiated fields are received by the radar, the time-domain response appears as damped sinusoids with its fundamental frequency related to the target dimensions and the soil property [4]. This resonance signature could provide a length estimation of the target.
A fully-polarimetric GPR radar provides a
complete two by two scattering matrix under the far-field assumption by
recording both co-polarization and cross-polarization data. This scattering matrix can be formed at a
given frequency or at a given time depending on whether frequency-domain or
time-domain data are used. In this paper, almost all processing use the
time-domain scattering matrix. The eigenvalues associated with this scattering
matrix are designated as 
and 
for the scattered components that are parallel and
perpendicular to the UXO axis, respectively. To characterize the linearity of
the target geometry, an estimated linearity factor (ELF) is defined as
. (1)
For a
rotationally symmetric (viewed from the incident waves) target, 
would be equal to
and ELF becomes minimum, i.e. zero. On the other hand, for a highly linear target such as a long and
thin conducting wire, 
would be negligible and ELF becomes maximum, i.e. one. Notice that, like the eigenvalues, the ELF
can either be a function of frequency or a function of time. The UXO’s azimuth
orientation an also be obtained from the eigenvectors associated with the
dominant eigenvalue, 
. Detailed discussion
about these features can be found in [3]
III. Technology Limitations
It was found that the major source of the missed UXO, .i.e. UXO’s classified as non-UXO’s, was the large inclination angle. When the sensor is located directly above an UXO, as adopted previously, both scattering strength and feature excitation degrade significantly when the inclination angle becomes greater than 45 degrees. This resulted in a lower rate of successful classification.
Investigation of the source of false alarms revealed that many false alarms observed in the previous test were related to small, shallow objects that were offset from the antenna center. Such an offset results in unbalanced illumination and shows a similar polarization property as a linear object like an UXO.
The above limitations occur because only a single location (directly above the hot spot) was used for features extraction. This situation can be improved by including features extracted from multiple locations as will be demonstrated in the next section.
The current technology is also unable to discriminate a true UXO from a non-UXO object that has an elongated body shapes. Several other possible features have been proposed and are currently under investigation [5][6].
IV. Technology Improvements
Base on the findings from the previous Tyndall test results. Several modifications have been made to improve the classification capability. These improvements are described below.
The previous antenna was orientated such that the antenna arms are at 45 degrees with respect to the vehicle heading as shown in Figure 1. This would work well if only features extracted with antenna directly above the target are utilized. A new orientation with the antenna arms parallel and transverse to the vehicle heading as shown in Figure 2 is required for the new multiple position scheme. This is due to the polarization property of the radiated field distribution that has a minimal depolarization level along the antenna arm directions. This is shown in Fig.3, a contour plot of the magnitude of the depolarized field.
Figure 1 Font-mounted antenna oriented 45 degrees with respect to the motion direction.
Figure 2 Towed antenna oriented parallel/transverse to the motion direction.
Figure 3 Calculated magnitude of depolarized fields near ground surface.
· New Measurement Approach
During the Tyndall test, a two-step scanning approach was adopted. During the first step, i.e. initial scan, multiple-position (5~7 positions) GPR data were collected along an arbitrarily oriented line passing through the target location. If the extracted features indicated UXO-like features with good linearity, two additional passes were made in the directions parallel and transverse to the target orientation estimated from the initial pass. This was designed to improve on the data from the original pass by measuring again with antenna arms closely aligned with the principal target axis. On the other hand, if the initially extracted features did not show good linearity, we surmised that further passes were unlikely to improve upon that. The target was then treated as a non-UXO item and no further action was taken. It was later found that a large percentage of missed UXO’s had only single-pass data. The problem lay in the tendency of a UXO with large inclination angle to produce poor linearity data features when the antenna is located directly above. Therefore, such UXO’s were excluded after the first pass.
The new approach, a minimum of two orthogonal
passes will be performed on each target during the first step scan. The
direction of the additional pass for a target that does not show linearity
during the initial pass will be orthogonal to the direction of the initial
pass.
Figure 4 plots the estimate linear factor (ELF) feature [2] extracted from the measured data for a conducting sphere shallowly buried in a sand as the radar scanned across it using the new antenna configuration. The spacing between the measurements was 3 inches. As one can see, the low ELF values are observed when the antenna is above the sphere as they should be for a non-linear object. As the position offset from the sphere increases, the ELF rises significantly. Such an increase of the ELF is because the sphere becomes physically closer to the antenna’s parallel element than the transverse element and thus, generating misleading ELF that resembles a linear object. Using the previous approach and exploiting data features obtained at an offset position, this sphere would have been classified as an UXO-like target due to the high ELF. This was indeed the case for many false alarms observed during the Tyndall results.
If one utilizes the spatial behavior of the ELF
as shown in Figure 4, one should be able to avoid this type of false alarm
by noting the dip near the center position.
For instance, the ELF and the estimated target orientation (ETO)
features for a horizontal conducting pipe 18 inches in length and 1 inch in
diameter are also shown in Figure 5 as a comparison. The pipe was oriented at 90 degrees. It is observed that the ELF remains high
over the scan range (
36 inches). Should a longer scan
have been taken, these feature parameters would have decreased to levels
corresponding to no-target area. The
extracted ETO also shows approximately 90 degrees, as it should be, over the
whole range. Comparing the spatial
behavior of the extracted features from the pipe with those from a sphere (a
non-linear object), the difference is quite evident. It is expected that most horizontal UXO’s will have similar
behavior as the pipe.
Another difficulty encountered during the Tyndall test was to discriminate a UXO with an inclination angle greater than 30 degrees. Figure 6 shows ELF extracted from the same pipe but with a 45-degree inclination angle, as shown in the illustration. The distance between the antenna aperture and the highest point of the pipe is approximately 15 inches. A large ELF value is observed for in most region as expected for a linear target except that the value begin to degrade when the antenna is offset to the left more then 15 inches. ELF reaches minimal value at approximately 20-inches offset to the left of the target. It is interesting to find that this is when the antenna happens to be looking into the end of the pipe. Therefore the degradation of ELF is not surprising since no resonance is excited at this incident angle. Similar results were also obtained for the same pipe at 60-degree and 90-degree inclination angles and are shown in Figure 7 and Figure 8, respectively. The distances between the antenna aperture and the highest point of the 60-degree and 90-degree pipe are approximately 13 and 11 inches, respectively. As in the 45-degree pipe case, the linear features degrades when the antenna position is such that it "views" the pipe more or less along its axis. Notice that the spatial ELF feature of the vertical pipe is very similar to that of the sphere shown in Figure 4 near the center. Other means such as resonance and scattering pattern have to be incorporated in this case for discrimination.
From the examples given above, it is evident that the spatial variation can improve discrimination capability. Robust algorithms that are required to characterize the spatial variations of the ELF, DEN and ETO are currently being developed.
Figure 4 The ELF extracted at different antenna positions for a 3-inch conducting sphere shallowly buried.
(a) ELF vs. antenna position.
(b)
ETO vs. antenna position.
Figure 5 The ELF and ETO extracted at different antenna positions for an 18-inch long, 1-inch diameter conducting pipe with an orientation transverse to the scan direction.
Figure 6 ELF for an 18-inch long, 1-inch diameter conducting pipe with an orientation parallel to the scan direction. The inclination angle was 45 degrees (see illustration).
Figure 7 ELF for an 18-inch long, 1-inch diameter conducting pipe with an orientation parallel to the scan direction. The inclination angle was 60 degrees.
Figure 8 ELF for an 18-inch long, 1-inch diameter conducting pipe with a inclination angle of 90 degrees.
V. Conclusions
The UXO classification technology based on broadband, fully-polarimetric radar feature were briefly reviewed. Important improvements in system configuration, measurement approach and processing technique that could lead to improved classification capability were discussed. These improvements were motivated by extensive study of the missed UXO’s and false alarms resulted from the previous field test that containing a challenging variety of target set. Preliminary measurement results also indicated that the additional spatial feature can indeed improve the UXO classification capability. In particular, many false alarms resulted from position offset could now be avoided. Good resonance and linear feature can be obtained for UXO’s with moderate inclination angles up to 60 degrees. Classification of vertical UXO’s are still quire challenging and require additional features.
VI. Acknowledgements
This work and its publication were supported by the United States Department of Defense's Environmental Security Testing and Certification Program (ESTCP).
VII. References
[1]
C.-C.
Chen and L. Peters Jr., "Buried Unexploded Ordnance Identification via
Complex Natural Resonances", IEEE
Trans. on Antenna and Propagations, vol. 42, pp. 1645-1654, Nov. 1997.
[2]
C.-C.
Chen, M.B. Higgins, K. O’Neill and R.
Detsch, “Tyndall AFB Measurement Results of UXO Characterization Using
Full-Polarimetric Ultra-Wide Bandwidth (UWB) GPR”, Proc. UXO/Countermine Forum, Anaheim, California, May, 2000.
[3]
C-C. Chen, M.B. Higgins, K. O'Neill and R. Detsch,
"UWB Fully-polarimetric GPR Classification of Subsurface Unexploded
Ordnance", Accepted for publication at IEEE Transaction on Geoscience
and Remote Sensing.
[4]
C.E.
Baum, “The SEM Representation of Scattering From Perfectly Conducting Targets
in Simple Lossy Media,” Phillips
Laboratory Interaction Note 492,
Apr. 1993.
[5]
K.
O’Neill, “Discrimination of UXO in Soil Using Broadband Polarimetric GPR
Backscatter,” IEEE Trans. Geosci. Remote
Sensing, in press, 2000.
[6]
J.D.
Young, “Radar Imaging from Ramp Response Signatures”, IEEE Trans. on Antenna and Propagations, vol. 24, pp. 1645-1654,
May 1976.