IEEE - Aerospace and Electronic Systems - March 2021 - 57

Image credit: Image Licensed by Ingram Publishing

b) Opposed to existing surveys [26]-[32], we specifically focus on the MSTAR dataset rather than presenting solutions involving other SAR datasets.
This is important as the MSTAR has established
itself as the major benchmarking dataset in the SAR
ATR domain.
c) Despite the numerous papers exploiting the MSTAR
database (see Figure 1), this article covers the most
recent and advanced progress of SAR ATR. Therefore, it provides the reader with state-of-the-art
methods.
d) Comprehensive comparisons of existing ATR algorithms with summaries, performance assessments,
and future research directions are also presented.
The structure of this article is as follows. We present
the MSTAR dataset in the next section, followed by the
various SAR ATR methods. The section " Performance
Analysis " compares and contrasts the performance of the
presented ATR algorithms. Last, we conclude the article
and propose future research directions.

MOVING AND STATIONARY TARGET ACQUISITION AND
RECOGNITION DATASET
The MSTAR dataset comprises X-band SAR imagery with
a 1x1 ft resolution, 360 articulation with a 1 spacing,
and an image size of 128 x 128 pixels [32]. The commonly
used targets in the literature are presented in Table 2 and
Figure 2, while for a complete list of all target variants
including the SLICY static target, the reader is referred to
[23]. For completeness, SLICY is a precisely designed
and machined engineering test target containing standard
radar reflector primitive shapes such as flat plates, dihedrals, trihedrals, and top hats. The purpose of this target is
to allow Image Understanding developers the ability to
validate the functionality of their algorithm with a simple
known target. However, it should be noted that despite the
MSTAR dataset offers SLICY, the majority of the literature does not involve it during SAR ATR evaluation and
thus in this article, it will be omitted.
The proposed SAR ATR methods are evaluated on
two major target configuration sets namely the standard
MARCH 2021

operation conditions (SOC) and the extended operation
conditions (EOC). The former typically includes the ten
target classes of Table 2 or a subset comprising of three
targets. EOC involves the same ten-class targets but with
various acquisition geometries, i.e., depression angle variations up to 45 , target state variations, i.e., articulation,
alternative configurations, and intraclass variability, and
finally local target deployment, i.e., various obscuration
levels [33], [34].
However, despite the MSTAR dataset has established
itself as the major SAR ATR benchmarking dataset, it is
worth noting that it suffers from various deficiencies that
can be grouped into dataset-linked and user-linked. Datasetlinked deficiencies include background correlation issues,
and a priori known training and testing imagery, while userlinked deficiencies a variable target patch size extracted
from the entire SAR image, and nonstandard SOC and EOC
trials. An analysis of these shortages involves
a) Background correlation. This is one of the major
dataset-linked flaws and is related to the strong
background correlation among the SAR images
allowing high true positive ATR rates even if the
target is absent from the scene [35], [36].
b) Nonstandard target patch size. A number of papers
aim to reduce the effectiveness of the background
correlation either by cropping the target from the
scene using a fixed mask size or by exploiting a target segmentation strategy. However, the final SAR
patch size that contains the target is not fixed and
thus may still contain background regions affecting
the final ATR performance, and making a direct
comparison between the available ATR methods
not trivial.
c) Nonstandard SOC and EOC trials. Despite [33],
[34] define the SOC and EOC target setups, over the
years additional SOC and EOC setups have
emerged by establishing various interclass (different
classes of targets) and intraclass target configurations (variations of the same target class) from the
standard MSTAR targets presented in Table 2.
However, the multiple SOC and EOC setups prohibit an extended and direct comparison among the
proposed SAR ATR algorithms, and evaluations are

IEEE A&E SYSTEMS MAGAZINE

IEEE - Aerospace and Electronic Systems - March 2021

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