IEEE Geoscience and Remote Sensing Magazine - December 2021 - 179

The image is strongly cluttered with contributions from the
rough surface.
The aforementioned tomographic imaging methods are
based on free-space approximation, neglecting the presence
of the air-to-ground interface and assuming the propagation
as occurring in a homogeneous dielectric medium. The performance
of the approximate free-space tomographic imaging
was contrasted in [55] with that of a tomographic algorithm
that accounts for the presence of the actual half-space
geometry. The latter implements the spectral representation
of the dyadic Green's function. Using numerical electromagnetic
FL-GPR data, the authors in [55] demonstrated that a
free-space approximation can lead to a loss of imaging resolution
and degradation in the SCR, as compared to its halfspace
counterpart. The impact of the lower resolution was
also observed in the estimated target statistics [53].
DATA-ADAPTIVE AND COMPRESSIVE
SENSING METHODS
A data-adaptive approach for FL-GPR image formation was
proposed in [44] (Figure 6). It is based on amplitude and phase
estimation and rank-deficient robust Capon beamforming.
There were 12 evenly spaced scans (each scan covering 2 m
in the down range) used to form the entire image, covering
24 m in total. The amplitude- and phase-estimation algorithm
in conjunction with the robust Capon beamformer provided a
significantly enhanced image quality compared to BPA.
Compressive sensing (CS) methods can also be applied
to exploit the intrinsic sparsity of the illuminated
scene in terms of the number of buried targets. A CS approach
was employed in [56] for scene reconstruction using
measurements from a MIMO FL-GPR system. Assuming a
linear model relating the measured data and the unknown
scene reflectivity, the image formation can be posed as a
solution to an inverse problem regularized by a sparsity-inducing
norm. This framework permits scene reconstruction
with spatial and temporal sampling at sub-Nyquist rates.
In real environments, even with few targets, there
exists strong clutter that populates and subsequently
degrades the quality of the reconstructed image. This is
because the rough surface clutter in the FL-GPR can be
distributed over the entire region. An FL-GPR image from
[56], generated by processing real data of a shallow-buried
metallic antitank landmine using the CS technique, is depicted
in Figure 7. Clearly, without clutter suppression, it
-1
1
2
3
40
45
50
Down Range (m)
(a)
-1
1
2
3
40
45
50
Down Range (m)
(a)
FIGURE 6. Real-data-based, single-look imaging results: (a) a
BPA imaging result and (b) the results of a hybrid of amplitudeand
phase-estimation algorithm and robust Capon beamformer.
(Source: [44].)
1
2
3
4
5
6
P10
-1
P10
P10
0123456 7
Y Distance (m)
FIGURE 5. An FL-GPR image of plastic mines. The crosses point to
the nominal buried locations of the targets. The label P10 denotes
a plastic mine buried at a depth of 10 cm. As expected, the image
is strongly cluttered by the contribution from the rough surface.
The blank region indicates where a strong stake (fiducial) return
has been masked from the image to make the mine returns more
visible. (Source: [30].)
DECEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
15.4 -2
-1
1
2
3
15.6 15.8 16 16.2
-0.8
-0.6
-0.4
55
60
-5
-10
-15
55
60
-5
-10
-15
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
14 14.5 15 15.5 16 16.5 17 17.5 18
Range (m)
FIGURE 7. A CS image of sparse data without clutter suppression.
(Source: [56].)
179
X Distance (m)
Cross Range (m)
Azimuth (m)
Cross Range (m)
(dB)
(dB)
IEEE Geoscience and Remote Sensing Magazine - December 2021

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2021
