IEEE Geoscience and Remote Sensing Magazine - March 2021 - 140

Detector

Search Regions
of Interest

Discriminator

Reject Natural
False Alarms

Classifier

Filter Human-Made
Objects; Predict Target Type

FIGURE 1. The backbone of three-phase automatic target recognition.

to detect and identify mobile targets (less than 1 × 1 ft) in
a variety of landscapes and deployments [2]. Meanwhile,
the data taken from spaceborne (Lacrosse and Discover satellites) and airborne (U2 reconnaissance plane and Global
Hawk unmanned aerial veSUPPOSE THE MISSIONS OF
hicle) platforms are too many
TARGET DETECTION AND
to be interpreted in a timely
CLUTTER DISCRIMINATION
manner. There is a pressing
ARE ALREADY OVER. GIVEN
need for automatic image
interpretation [3]-[6]. A typiTHE CHIP IMAGES OF A
cal
automatic target recogniCERTAIN TARGET, OUR TASK
tion
system usually follows a
IS TO PREDICT THE
three-phase workflow of tarIDENTIFICATION.
get detection, clutter discrimination, and classification, as
shown in Figure 1. The input
is a scenery of SAR images containing several potential targets, while the outputs are the types of target lists.
In the first phase, regions of interest (ROIs), including
targets and false alarms, are roughly located. The natural
false alarms include trees and stones, while human-made
false alarms refer to buildings, corner reflectors, and streetlights. The most popularly used approach to target detection is the constant false alarm rate algorithm (CFAR) [7],
[8]. The core idea is to calculate the threshold according to
the false alarm rate and statistical property of clutter under
the premise of a constant false alarm. The representative is a
two-parameter CFAR detector. It identifies potential targets
by searching for the bright returns. The amplitude of the test
cell is compared with the mean and standard deviation of
the cells around it:

TABLE 1. THE BASIC VERSIONS OF A CFAR DETECTOR.
CFAR

DISTRIBUTION

CA-CFAR/Cell-averaging CFAR [9]

Gaussian [2]

GO-CFAR/Greatest of CFAR [10]

Weibull [11]

SO-CFAR/Smallest of CFAR

K distribution [12]

OS-CFAR/Order-statistic CFAR [13]

a -stable distribution [14]

140

x q - nt
2 TCFAR & target, (1)
#
vt
clutter

where x q is the amplitude of the test cell, nt is the estimated
mean of the clutter, vt is the estimated standard deviation
of the clutter, and TCFAR is a constant threshold value resulting from the false alarm rate. (Although the clutter in highresolution SAR images does not follow a Gaussian distribution, this still works well for detecting targets in the clutter.)
Some improved versions are listed in Table 1.
In the second phase, the task is to determine whether
the potential target pixels (ROIs) come from a true target
and eliminate the naturally occupying clutter. The oneclass discriminator is developed to provide each ROI a score
that measures the distance from the candidate to the center
of the training set of the target in a feature space, such as the
quadratic discriminator [15]:

1
t hT Ct -1 ^X i - M
th
d i = n ^X i - M

for i = 1, 2, f, n t + n c,

(2)

t and
where n is the number of discrimination features; M
t
C are the estimated mean vector and variance-covariance
matrix of target, respectively; X represents the observed
candidate feature; and d i is the distance from the test to
the target class. The variables n t and n c are the number of
targets and the number of clutter false alarms, respectively.
The lower the distance, the more target-like the candidate.
The choice of features, therefore, plays an important role for
clutter discrimination. Some representative discriminating
features are summarized in Table 2.
In the third phase, a classifier is trained to reject the cultural-clutter false alarms and predict the types of targets on
the list. It goes into the range of pattern recognition. Various machine learning tricks are introduced to reach the inference. The typical decision engines include k-NN, support
vector machine (SVM), boosting, decision tree, and regression models [20]-[25].
This article pays more attention to the third phase, classification. Suppose the missions of target detection and clutter discrimination are already over. Given the chip images
of a certain target, our task is to predict the identification.
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

MARCH 2021

IEEE Geoscience and Remote Sensing Magazine - March 2021

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