IEEE - Aerospace and Electronic Systems - March 2021 - 58

Automatic Target Recognition on Synthetic Aperture Radar Imagery: A Survey
Table 1.

Analysis of Data Domain-Driven ATR Solutions
3D LIDAR

2D
visual

2D IR

2D SAR

single 3D
model

multiple
views

Target pose
invariance

Target
illumination
invariance

ATR based
on the
underlying
structure

Operating
day and
night

ATR
Processing
time

Equipment
cost

Equipment
size

Power
consumption

Data
acquisition
rate

(scanning
LIDAR)
þ (flash
LIDAR)

Maximum
operating
range

%100 m

>100 m

<100 m

>>100 m

active

passive

active

Template
size

Figure 1.
MSTAR citation progression.

constrained to common target subsets. In fact, the
MSTAR related literature offers more than 17 types
of 10-class and 3-class SOC trials, and more than 33
types of EOC.
d) A priori known training and testing imagery. Both
training and testing target sets are a priori known
making the evaluation unrealistic and questioning
the true effectiveness of the proposed methods
against unknown testing sets.
Though, despite these deficiencies, the MSTAR dataset is currently the standard SAR ATR benchmarking
dataset with numerous citations (see Figure 1). The most
common SOC and EOC configurations presented in the
literature are shown in Table 3. However, for better readability, the performance comparison of various current
techniques presented in the conclusion will involve only
the most common MSTAR evaluation subsets presented
in Table 3.

ENCODING AND CLASSIFYING SAR IMAGERY
SAR image encoding, i.e., a compact tensor-type of representation either describing the SAR image on a local or a
global neighborhood basis, is the first major stage for a
SAR ATR system. Conceptually, the simplest approach is
to neglect any data domain remapping and directly encode
the raw SAR reflectivity. Though, the simplicity of this
strategy does not underestimate the complexity of the
encoding process itself but rather reflects the immediacy
of applying the encoding technique. On the other side,
some methods initially transform the SAR data into a different data domain and then encode it. Hence, we define
two major encoding schemes, i.e., the reflectivity attribute-based, and the transform-based, accordingly. However, regardless of the encoding scheme and data domain
employed, the second major phase of ATR incorporates
matching the template encodings against the target one
and ultimately classifying the target. Despite the encoding
and classification phases are discrete, the SAR ATR methods offered by the literature are presented as a complete
algorithm where the two major phases are heavily associated and cross-tuned to gain the maximum ATR performance. Thus, it is more appropriate to present and
evaluate each technique as a whole pipeline rather than
58

Reveal
sensor
position

two isolated schemes. Hence, for better readability, we
group the current SAR ATR methods based on the data
domain utilized and subgroup them depending on the governing method employed, i.e., the technique mostly affecting the classification performance among the ones used in
a specific SAR ATR architecture.

REFLECTIVITY ATTRIBUTE-BASED METHODS
This type of technique exploits the raw SAR imagery
without involving any preprocessing phase to alter the
initial data domain. As described in the following subsections, this group of methods involves, feature-based
methods, scattering centers calculation, low-rank
matrix factorization, deep learning, sparse representation classification, and hybrid strategies combining two
of the reflectivity methods.

IEEE A&E SYSTEMS MAGAZINE

MARCH 2021

IEEE - Aerospace and Electronic Systems - March 2021

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - March 2021