IEEE Geoscience and Remote Sensing Magazine - March 2021 - 141

MAIN CHALLENGES
Although a great many works have been initiated previously to provide a baseline knowledge of target characteristics, the problem is still far from being solved due to the
complex application scenarios and the mutable target scattering phenomenology. The Moving and Stationary Target
Recognition (MSTAR) program evaluates the recognition
performance against a set of extended operating conditions
(EOCs) [26], where the operational parameters between the
images used for training and those available for testing
are different.
In addition to the full 360° (0°-359°) target aspect view
coverage over a range of depression angles, the EOCs include squint angle changes, target articulation and configurations, structural modifications, obscuration resulting
from the occlusion and/or layover, and intraclass variability.
The main challenge of target recognition can, therefore, be
summarized as follows.
◗◗ Aspect and elevation angle variations: A peculiar characteristic of SAR images is that even a small change in aspect
relative to the incident energy direction alters the target
signatures significantly. The aspect sensitivity is, therefore, much more intense than in the optical imagery due
to the following two factors:
1) the interaction between the incident electromagnetic
wave and the target's physical structure
2) the coherent superposition of returns from different
positions of the target structure.
Figure 2 draws a set of the signatures of BMP2 with a
1° incremental aspect view. The stable pixels result from
the strong backscatter returns, while the varying pixels
are due to the changes in phasing between individual
scattering cells of comparable return strength.
◗◗ Similar target signature: The goal of target recognition is to
distinguish among targets by exploiting the differences
in their individual signatures. The differences are subtle and localized. Figure 3 shows the signatures of the
BTR60 and BTR70 (armored personnel carriers) with
five aspect views. The sets of images exhibit very similar
target signatures. Conversely, the outliers and the local
regions behave obviously differently. It is these pixels
and regions that carry the vital information that can be
used to distinguish these two tanks.

(a)

(b)

(c)

(d)

TABLE 2. THE REPRESENTATIVE FEATURES
FOR TARGET DISCRIMINATION.
DEVELOPER

FEATURES

Lincolna

Standard deviation, fractal dimension, and weighted-rank
fill ratio
Size-related: mass, diameter, and square-normalized
rotational inertia

ERIMb

Contrast-based: MaxCFAR, mean CFAR, and percent
bright CFAR
Polarimetric: percent pure, percent pure even, and
percent bright even

Rockwellc

Polarimetric: phase ratio and specific entropy

Lorald

Contiguousness

Principle

The first- and second-order moment of local image
amplitude [16], [17]

Bhanu

Projection: horizontal, vertical, and minor and major
diagonal projection
Distance: minimum, maximum, and average distance [18]
Moment: horizontal, vertical, and diagonal second-order
moments

NUDT e

Average and peak signal-to-noise-ratio and percentage of
bright pixels [19]

a These

features are developed under the Strategic Target Algorithm Research contract
funded jointly by the Advanced Research Projects Agency and the U.S. Air Force and
administered by Lincoln Laboratory.
b Environmental Research Institute of Michigan, Ann Arbor.
c Rockwell International Corporation of Goodyear, Arizona.
d Loral defense system.
e National University of Defense Technology, Changsha, China.

◗◗ Depression angle changes: Different from the optical sen-

sor, the SAR image plane into which the backscattered returns are projected is determined by the squint
angle and the depression angle and, hence, contains
the line of sight of the radar. Changing the depression
angle impacts both the projection of signature returns
and the target shadow into the SAR plane. Figure 4
shows the signatures of 2S1 across four depression
angles. The signatures stretch along the range with
the depression angle increased. The variations in the
composite pixel brightness result in superimposition
due to changes in the coherent interference of the different returns.
◗◗ Squint angle changes: The SAR image plane results from
the squint angle and the depression angle. Changing the

(e)

(f)

(g)

(h)

FIGURE 2. The SAR images of a BMP2 infantry fighting vehicle with (a) 23°, (b) 24°, (c) 25°, (d) 26°, (e) 27°, (f) 28°, (g) 29°, and (h) 30°
aspect views taken at a 15° depression angle. Even a change in the aspect view of an increment of 1° alters the scattering phenomenology
significantly.
MARCH 2021

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

141

IEEE Geoscience and Remote Sensing Magazine - March 2021

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