http://www.mrw.interscience.wiley.com/erfme/articles/eme347/abstract.

html

Radar Target Recognition

Kie B. Eom, The George Washington University, Washington, DC

Encyclopedia of RF and Microwave Engineering

Copyright © 2005 John Wiley & Sons, Inc. All rights reserved.
DOI: 10.1002/0471654507.eme347
Article Online Posting Date: April 15, 2005

Keywords: radar target recognition; sensor fusion; hierarchical modeling; neural network

Abstract
Radar target recognition technology has advanced remarkably as radar technology advances recently.
The resolution and features of modern radar system provide enough information to recognize tactical
targets. Different types of radar, such as synthetic aperture radar (SAR ), milimeter-wave (MMW ) real
aperture radar (RAR ), and interferometric synthetic aperture radar (IFSAR ), have been explored for
radar target recognition. Target recognition from radar signatures is challenging since the radar signature
is very sensitive to changes in look angle. There have been different approaches utilizing unique
characteristics of different radar types, and they are reviewed in this article. They are hierarchical
modeling of MMW RAR radar signatures for noncoorpertative target recognition, fusion of IFSAR data
Earth orbit data for discriminating targets of different heights, the neural network approach, spatial
matched filter for SAR , and fusion of multiple sensors.

About Wiley InterScience | About Wiley | Privacy | Terms & Conditions

Copyright© 1999-2005 John Wiley & Sons, Inc. All rights reserved.

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/abstract.html22.08.2005 16:09:35
 Introduction

Radar Target Recognition

Kie B. Eom, The George Washington University, Washington, DC

Encyclopedia of RF and Microwave Engineering

Copyright © 2005 John Wiley & Sons, Inc. All rights reserved.
DOI: 10.1002/0471654507.eme347
Article Online Posting Date: April 15, 2005

1. Introduction
Recent (as of 2003) advances in radar provides sufficient resolution and enough information to
recognize tactical targets from radar returns (21). For target recognition applications, different types of
radar, such as synthetic aperture radar (SAR), millimeter-wave (MMW) real aperture radar (RAR), and
interferometric synthetic aperture radar (IFSAR), have been explored (1-4). Modern radar system also
provides extensive data including fully polarimetric and Doppler channels (14, 26, 32), but there still are
many challenges for target recognition using radar returns (16, 17). This is because special
characteristics of radar and recent advances of radar technology make the recognition of targets difficult.
For example, the radar profile changes drastically for the small change in viewing angle, and the recent
development of stealth technology significantly alters the radar signature. In radar target recognition,
many different approaches for different radar types are suggested, and are discussed in this article.

Synthetic aperture radar provides high resolutions for both range and azimuth directions, and have been
used widely for target recognition applications. Since SAR is an imaging radar, it can provide detailed
images of target area with cloud-penetrating characteristics. When its fully polarimetric data is utilized,
targets can be recognized with high accuracy. The SAR target recognition is typically done in multiple
stages (18). The first stage of a typical SAR target recognition algorithm is a prescreener, where regions
of interest are located in this stage. Constant false-alarm rate (CFAR) detection algorithm is often used
in the first stage. The second stage may be a discriminator, where natural clutter false alarms are
rejected. In this stage, textural features are extracted from the target-size windows applied to the
neighborhood of pixels located by CFAR detection. The third stage is typically a classifier, where the
targets are classified and manmade discretes are rejected. Extensive studies in statistical pattern
recognition have been done, and many different pattern classifiers are used in target recognition
applications. For SAR target recognition, spatial matched filters have been investigated (18) for the
recognition of ground vehicles. Novak's (18) spatial matched filter approach in target recognition is
explained in later sections.

Although SAR technology provides many advantages over real aperture radar (RAR), there are many
applications where RAR is important. The azimuth resolution of SAR is much better than that of RAR,
but it is typically based on the assumption that the movement of the radar is at constant speed and
direction in SAR processing. When the radar is moving at rapidly changing speed and trajectory, as in
the case of the radar mounted at fighter airplanes or self-guidance weapons, the assumptions for SAR

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect1.html (1 of 3)22.08.2005 16:09:52

 Introduction

processing are not correct. Millimeter-wave (MMW) RAR represents the next generation of military
smart sensors for detection, tracking, and surveillance because of its high-range resolution, which is
critical in target recognition (6). The MMW RAR technology is sufficiently advanced, and the range
resolution of MMW radar is sufficiently high to discriminate tactical targets at a distance of several
kilometers (31). The target recognition by hierarchical modeling of high-range resolution (HRR)
millimeter-wave (MMW) radar signatures is discussed in this article.

The artificial neural network (ANN) has been widely used in target recognition. The use of ANN in
radar target recognition is also discussed in this article. ANNs are used as pattern classifiers, feature
extractors, model adaptation, fuzzy classification, and in other applications in radar target classification.
Feedforward neural networks have been used as pattern classifiers after the training with known target
samples. The learning algorithm of neural networks provides a powerful tool for adaptation of the
classifier to input vectors (24). A self-organizing feature map has been used as a feature extractor (27)
for radar target recognition. In combination with Kohonen's learning vector quantizer (LVQ) for
supervised classification, it has been applied to the recognition of ground vehicles from MMW HRR
radar signatures (27). Perlovsky et al. (19) suggested that the model-based neural network to include a
priori information to an ANN. This approach can reduce the search space of the neural network by
incorporating a priori information to the adaptability of an ANN. The fuzzy neural network (FNN)
approach (25) is also suggested to classify targets that may belong to more than one class. The advances
in neural network approaches can potentially improve the performance of target recognition algorithms
further. There are many approaches to incorporate information from many different sources for radar
target recognition. By fusing the information from more than one sensor, the accuracy of radar target
recognition may be improved. Two approaches in utilizing information from multiple sensors are
discussed in this article. Interferometric synthetic aperture radar (IFSAR) provides elevation
information, in addition to two-dimensional radar image, by processing interference between radar
returns received by two different antennas. By processing IFSAR image and fusing to SAR of visual
images, the accuracy of the target recognition can be improved substantially. The approach of
combining IFSAR and visual images using the image registration approach is discussed in this article.
There are statistical approaches in data fusion, and Bayesian data fusion approaches are used in radar
target recognition (13). In this approach, features from polarimetric SAR images are fused to improve
the recognition accuracy.

Radar target recognition is a complex problem, and no single algorithm performs better than other
algorithms with different types and modes of radar. In this article, different approaches for radar target
recognition is discussed in terms of radar types and approaches.

About Wiley InterScience | About Wiley | Privacy | Terms & Conditions

Copyright© 1999-2005 John Wiley & Sons, Inc. All rights reserved.

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect1.html (2 of 3)22.08.2005 16:09:52

Introduction

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect1.html (3 of 3)22.08.2005 16:09:52

 Hierarchical Modeling Approach for Noncooperative Target Recognition

Radar Target Recognition

Kie B. Eom, The George Washington University, Washington, DC

Encyclopedia of RF and Microwave Engineering

Copyright © 2005 John Wiley & Sons, Inc. All rights reserved.
DOI: 10.1002/0471654507.eme347
Article Online Posting Date: April 15, 2005

2. Hierarchical Modeling Approach for Noncooperative Target Recognition

The target recognition using hierarchical modeling of MMW RAR radar signatures is considered in Ref.
10, and the algorithm is tested with MMW radar data from noncooperative target identification (NCTI)
database. In radar target recognition problems, the targets may be shifted and scaled from the trained
data. The change in viewing angle results scale change in radar signature, and the change in azimuth
angle change results shifting in radar signature. Therefore, a classifier using features extracted from the
radar signatures should classify targets reliably even when radar signatures are shifted or scaled. In
general, parameters of time-series models such as ARMA (autoregressive moving-average) are not scale-
invariant (20, 22). Thus, the classification of scaled targets with model parameter features can result in
poor classification accuracy. There are a few approaches to make a classifier more reliable to shifting
and scaling. For example, the use of scale-invariant features or a training with scaled samples can make
a classifier more robust to scaling.

Hierarchical ARMA modeling approach is briefly discussed as follows. Suppose that a continuous signal
x(t) is a training sample. The classifier need to classify scaled signal of x(t) correctly as the same class as
x(t). The scaled signal of x(t) is given by

(1)

One approach to achieve this is to train a classifier with features extracted from scaled signals of x(t).
For example, different features at m different scales are extracted from scaled signals
, then the classifier is trained with these multiscale features. If the number of
scales included in the training is large enough, the classifier will classify signals having large scale
changes.

However, there are at least two potential problems with this approach if the signal is a discrete signal {x
(i), i=1, ¼, N}: (1) the original signal is defined only at discrete points—the signal at the finer scale is
not defined at certain points; and (2) feature extraction is performed multiple times with a single training
sample, and the computational complexity increases linearly as the number of scales increases. These
difficulties can be solved by the hierarchical modeling approach. The hierarchical modeling approach
presented in this section extracts multiscale features without adding much computational complexity.

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect2.html (1 of 7)22.08.2005 16:10:20

 Hierarchical Modeling Approach for Noncooperative Target Recognition

A discrete signal can be scaled to a coarser scale or a finer scale by decimation filtering or interpolation,
respectively (7, 8, 11, 12). We will first consider the decimation filtering of a signal, and its effect to the
statistical model, and then we will consider the scaling to a finer scale as a modeling process. A
decimation filter is defined as a local averaging (FIR filtering) followed by a downsampling process as
shown in Fig. 1 (29, 35). If the downsampling rate is m, the decimation-filtered signal represents the
signal at the scale reduced by a factor of m.

Figure 1. A decimation filter.

[Full View]

Let H be a FIR filter of length r and c be the downsampling operator of factor m:

(2)

(3)

Suppose that a signal at a coarser scale y (i) is obtained by decimation filtering of the original signal x
m

(i):

(4)

Suppose that the signal {x(i), i=1, ¼,N} follows an ARMA(p,q) model

(5)

where {w(i)} is a zero mean white-noise sequence with variance and the a and b terms are real
j j

coefficients. Equation (5) can be rewritten as

(6)

where

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect2.html (2 of 7)22.08.2005 16:10:20

 Hierarchical Modeling Approach for Noncooperative Target Recognition

(7)

and d is the unit delay operator, and we assume that the roots of A ( d) and B ( d) lie inside of the unit
p q

circle for stability and invertability of the model.

To find features at coarser scale, the model at a coarser scale should be considered. The following
theorem summarizes the results on the modeling of a decimation-filtered ARMA process.

*
Theorem 1. The decimation-filtered process {y (i)} defined in (4) follows an ARM A(p,q ) model,
m

where the order of AR polynomial is p and the order of the MA polynomial is q*=[(p(m–1)+r+q–1)/m],
and the model parameters can be obtained from the model parameters of x(i)

(8)

where

(9)

Also the AR polynomial C ( d) of aggregated data satisfies the following relation:

p

(10)

where are roots of A ( d).

p

The proof can be found in Ref. 10.

The ARMA model parameters are shift-invariant because the model's parameters depend only on mean
and correlations. Power spectral density can be estimated by using ARMA model parameters, and is also
shift-invariant and provides features that are intuitively appealing. For example, spectral peaks or
notches represents presence or absence of a frequency component in the signal. For radar signal
classification, ARMA power spectrum features at multiple scales are used. Power spectral density of an
ARMA process can be estimated by an extended least-squares (ELS) method.

Suppose that x(i) is an ARMA(p,q) process and w(i) is the input white sequence with variance s2 as

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect2.html (3 of 7)22.08.2005 16:10:20

 Hierarchical Modeling Approach for Noncooperative Target Recognition

defined in Eq. (5). Let R (k) be the autocorrelation functions of x(i), and R (k) be the cross-correlation
xx xw

between x(i) and w(i). The Yule–Walker equation for the ARMA process x(i) is given by the following
equation:

(11)

The AR (autoregressive) parameters are estimated by solving these Yule–Walker equations. By using
the estimated AR parameters, MA (moving-average) component of x(i) can be obtained by filtering AR
component from x(i):

(12)

The power spectral density of the ARMA process x(t) is estimated from the correlations of xma(t) and the
AR parameters estimated by Yule–Walker equations. The ELS power spectrum estimation algorithm is
summarized below.
ELS Spectrum Estimation Algorithm
Step 1. Compute sample correlations R (k) for k=0, ¼, p+q:
xx

(13)

Step 2. Estimate AR parameters a1, ¼,a by

p

(14)

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect2.html (4 of 7)22.08.2005 16:10:20

 Hierarchical Modeling Approach for Noncooperative Target Recognition

Step 3. Compute the sample correlation of MA component xma (i), which is obtained by removing
AR component:

(15)

Step 4. Compute the ARMA power spectrum:

(16)

For each training sample x(i), the models at the other scales (both coarser and finer scales) are obtained
by the hierarchical modeling approach presented in this section. The model at a coarser scale is obtained
using Theorem 1. The AR polynomial is obtained by Eq. (10), and the correlation of the signal at the
coarse scale is obtained with a proper choice of smoothing filter H, such as the Gaussian filter. Thus, the
spectral density of the signal at a coarser scale is obtained by the ELS algorithm. The model at a finer
scale is obtained by the approach explained in the next step sequence. The AR polynomial of the signal
at a finer scale are obtained under a no-hidden-periodicity assumption (34, 35). The correlation function
at a finer scale is obtained by disaggregation (10), and the ARMA spectrum at a finer scale is obtained
by the ELS estimation algorithm. The multiscale feature extraction algorithm is summarized below.
Multiscale Spectral Feature Extraction Algorithm
Step 1. Each radar return is normalized to zero mean and unit variance by

(17)

where and are sample mean and sample variances of x(i). M K-dimensional features from M
scales (including coarser and finer scales) are obtained from the normalized radar returns by the
following procedure.

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect2.html (5 of 7)22.08.2005 16:10:20

 Hierarchical Modeling Approach for Noncooperative Target Recognition

Step 2. For each training sample, the AR parameters and correlations are estimated by the ELS
algorithm. For k=0,1, ¼,K–1, the power spectrum is estimated at w= pk/K. The logarithm of the
power spectral density forms a K-dimensional feature vector.

Step 3. At each coarser scale, a feature vector is obtained by estimating the power spectrum using the
ELS method with model parameters obtained by the hierarchical modeling approach. The logarithm
of the power spectral density forms a K-dimensional feature vector at a coarser scale. Feature vectors
at multiple scales are obtained by repeating this step at coarser scales.

Step 4. At each finer scale, a feature vector is obtained by estimating the power spectrum using the
ELS method with model parameters obtained by the hierarchical modeling approach. This is repeated
for other finer scales, and multiple K-dimensional feature vectors are obtained from the logarithm of
the power spectral density.

Classification is done by a minimum-distance classifier with multiple prototypes. In this approach, each
training sample generates M prototypes corresponding to M scales. Therefore, if there are N training
signals for each class, then NM prototypes will be available for each class. Let us assume that there are
N prototypes in the class k∈{1, º,K}. For a test pattern x, the distance to the class k is
k

defined by

(18)

where the intersample distance d(x,z) is the Euclidian distance between x and z. The distance D is the
i

smallest of the distances between x and each prototype of the class k. The test pattern x is classified by
the minimum-distance decision rule: x is classified into class k if D < D for all i≠k.
k i

In Ref. 10, the hierarchical model-based features are tested with NCTI data. Figure 2 shows a typical
NCTI radar signature, and estimated power spectral density. In Ref. 10, about 95% of classification
accuracy is reported with 5000 MMW RAR radar signatures.

Figure 2. A HRR radar signature from NCTI database and its power spectrum estimated
by hierarchical modeling.
[Full View]

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect2.html (6 of 7)22.08.2005 16:10:20

Hierarchical Modeling Approach for Noncooperative Target Recognition

About Wiley InterScience | About Wiley | Privacy | Terms & Conditions

Copyright© 1999-2005 John Wiley & Sons, Inc. All rights reserved.

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect2.html (7 of 7)22.08.2005 16:10:20

 Neural Networks Approaches

Radar Target Recognition

Kie B. Eom, The George Washington University, Washington, DC

Encyclopedia of RF and Microwave Engineering

Copyright © 2005 John Wiley & Sons, Inc. All rights reserved.
DOI: 10.1002/0471654507.eme347
Article Online Posting Date: April 15, 2005

3. Neural Networks Approaches

Neural networks have been widely used in radar target recognition applications (19, 24, 25, 27). A good
survey on the use of neural networks in target recognition can be found in Ref. 24. Neural network
approaches have been popular because they have many promising qualities in target recognition
applications. Neural networks have the capability for adaptation to additional targets and environments.
The existence of powerful learning algorithms is one of the main strengths of the neural network
approach (24). They also offer techniques for selecting, developing, clustering, and compressing features
into a useful set, and provide automatic knowledge acquisition and integration techniques for target
recognition systems.

Feedforward neural networks have been used as pattern classifiers for target recognition (23). Let x be
i

the multidimensional feature vector extracted from a radar image, and let S be the index set of the target
patterns:

(19)

The optimum receiver (23) defined by

(20)

can be implemented as a neural network

(21)

where T is the threshold and Q(x) is the Heaviside step function. Roth (23) showed that detection of
target patterns out of a set of P of patterns can be handled by the feedforward neural network presented
above.

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect3.html (1 of 3)22.08.2005 16:10:25

 Neural Networks Approaches

Neural networks have been also used as a feature extractor for target recognition. Kohonen's self-
organizing map (SOM) and learning vector quantizer (LVQ) have been used in the two-stage target
recognition approach (27). SOM is based on unsupervised competitive learning where only one output
node, or one per local group of nodes at a time, gives the active response to the current input signal,
clusters input vectors into preselected C classes by adapting connection weights to nodes in the network,
and is used as a feature extractor (27). At each iteration of the SOM algorithm, the best matching node c
is selected by

(22)

where x is the current input vector and {m1, º, m } is the set of nodes (cluster centers). Then each node
C

m located in the neighborhood of the node c is adapted by the learning rule

i

(23)

where the gain h (t) can be a simple monotonically decreasing function of time or a Gaussian gain
i

function defined by

(24)

The learning rate a(t) and the kernel width s(t) are monotonically decreasing functions, and their exact
forms are not critical. LVQ is used as a supervised classifier of the features extracted by SOM (27).
Stewart et al. (27), reported greater than 94% accuracy in the target recognition experiment with MMW
data having five types of ground vehicles.

The model-based neural network (MBNN) was introduced (19) to combine a priori knowledge of
models of data with adaptivity to changing data properties. Both learning and adaptation of the MBNN
is done by iterative estimation of association weights and model parameters. Different statistical models
for different physical processes, background clutter, outlier, target pixels, and other elements are also
introduced in Ref. 19. This approach has a potential for improving target recognition performances by
allowing the inclusion of a priori information in addition to the adaptability of a neural network.

Fuzzy neural networks are also used in radar target recognition. Fuzzy ARTMAP and EMAP neural
networks are suggested (25) for radar target recognition. The fuzzy neural network allows one to make
soft decisions in classifying a target, and each input vector can belong to more than one class. The fuzzy
association between the input vector and the classified target can potentially improve the performance
and the complexity of the adaptation.

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect3.html (2 of 3)22.08.2005 16:10:25

Neural Networks Approaches

About Wiley InterScience | About Wiley | Privacy | Terms & Conditions

Copyright© 1999-2005 John Wiley & Sons, Inc. All rights reserved.

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect3.html (3 of 3)22.08.2005 16:10:25

 Exploitation of Elevation Data from IFSAR

Radar Target Recognition

Kie B. Eom, The George Washington University, Washington, DC

Encyclopedia of RF and Microwave Engineering

Copyright © 2005 John Wiley & Sons, Inc. All rights reserved.
DOI: 10.1002/0471654507.eme347
Article Online Posting Date: April 15, 2005

4. Exploitation of Elevation Data from IFSAR

IFSAR is a technique to generate high-resolution Digital Elevation Map (DEM) based on the phase
difference in SAR signals received by two spatially separated antennas (36). There are drawbacks in
height maps derived from IFSAR data: The data are noisy and the spatial resolution is far inferior to that
of visual data. The spatial resolution is further degraded by the noise removal step. Figure 3 shows a
height map produced by a real IFSAR. A typical IFSAR elevation image is noisy and needs to be filtered
before it can be used reliably. Also, there are regions with “no data,” resulting either from the fact that
the original scene was not on a rectangular grid or from radar geometry effects that cause some points
not to be mapped. Interpolation and nonlinear filtering techniques are used to filter the elevation data.

Figure 3. An IFSAR image.

[Full View]

Positioning of IFSAR and visual data allows for the fusion of clues from both sensors for target
recognition. It is needed to overcome various difficulties resulting from the limitations of the sensor. For
example, building detection requires the extraction and grouping of features such as lines, corners, and
building tops to form buildings (9). The features extracted from visual data usually include many
unwanted (spurious) edges and lines that do not correspond to buildings. The grouping stage requires
complex and computationally intensive operations. Further, the height of a building is typically
estimated by extracting shadows and sun angle when available, and is not reliable when the shadows are
cast on adjacent buildings. Another drawback of methods based exclusively on visual data lies in their
sensitivity to imaging conditions.

IFSAR elevation data can be used in conjunction with visual data to overcome the abovementioned
difficulties. Current IFSAR technology provides sufficient elevation resolution to discriminate building
regions from surrounding clutter. These building regions are not well defined from a visual image when

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect4.html (1 of 3)22.08.2005 16:10:27

 Exploitation of Elevation Data from IFSAR

the buildings have the same intensity level as their surrounding background. Similarly, a building having
different colors may be wrongly segmented into several buildings. IFSAR data are not affected by color
variations in buildings and therefore are better for building detection.

Figure 4 shows a visual image and edges detected by the Canny operator for the area shown in Fig. 3.
The left portion of Fig. 4 shows a building with two different roof colors and roof structures on many
buildings. Many spurious edges not corresponding to the building appear in the edge map shown on the
right of Fig. 4. Using the IFSAR elevation map shown in Fig. 3, buildings and ground regions are
labeled using a two-class classifier. The IFSAR and visual images are registered. Figure 5 shows the
result of registration of a visual image and the segmented elevation image. Features corresponding to
roads, parked cars, trees, and other objects are suppressed from the visual images using the segmented
buildings derived from the IFSAR image.

Figure 4. Visual image and edges detected by the Canny operator.

[Full View]

Figure 5. Buildings segmented from the IFSAR image overlaid to visual image.
[Full View]

The location and the directions of edges in the segmented image are estimated, and are used to locate
edges of buildings in the visual image. In the visual image, an edge pixel corresponding to each edge
pixel in the registered height image is searched in the direction perpendicular to the estimated direction
in the height image. If an edge is found within a small neighborhood, the edge pixel is accepted as a
valid edge of a building. If such a pixel is not found in the neighborhood, the edge is not accepted.
Figure 6 shows the refined edges obtained by searching in the neighborhoods of height edges. Most of
building edges in the height image are found while the unwanted edges are removed.

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect4.html (2 of 3)22.08.2005 16:10:27

Exploitation of Elevation Data from IFSAR

Figure 6. Edges Refined using IFSAR result.

[Full View]

About Wiley InterScience | About Wiley | Privacy | Terms & Conditions

Copyright© 1999-2005 John Wiley & Sons, Inc. All rights reserved.

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect4.html (3 of 3)22.08.2005 16:10:27

 Spatial Matched-Filter Classifiers for SAR

Radar Target Recognition

Kie B. Eom, The George Washington University, Washington, DC

Encyclopedia of RF and Microwave Engineering

Copyright © 2005 John Wiley & Sons, Inc. All rights reserved.
DOI: 10.1002/0471654507.eme347
Article Online Posting Date: April 15, 2005

5. Spatial Matched-Filter Classifiers for SAR

A typical target recognition using SAR is done in multiple stages, and is illustrated by the block diagram
in Fig. 7. In the first stage, a CFAR detector prescreens by locating potential targets on the basis of radar
amplitude. Since a single target may produce multiple detections, the CFAR detections are clustered
(grouped together). Then a region of interest (RoI) around the centroid of each cluster is passed to the
next stage of the algorithm for further processing.

Figure 7. Block diagram of a typical baseline target recognition system (adapted from
Novak (17)).
[Full View]

The second stage takes each RoI as its input and analyzes it. The goal of this discrimination stage is to
reject natural clutter false alarms while accepting real targets. This stage consists of three steps: (1)
determining the position and orientation of the detected object, (2) computing simple texture features,
and (3) combining the features into a discrimination statistic that measures how “targetlike” the
detection object is.

The third stage is classification, where a 2D pattern-matching algorithm is used to (1) reject clutter false
alarms caused by manmade clutter discretes (buildings, bridges, etc.) and (2) classify the remaining
detected objects. Those detected objects that pass the second stage are matched against stored reference
templates of targets. If none of the matches exceeds a minimum required score, the detected object is
classified as clutter; otherwise, the detected object is assigned to the class with the highest match score.

Matched filters are investigated in (18) as pattern-matching classifiers in target recognition system
shown in Fig. 7. They are the synthetic discriminant function (SDF), the minimum average correlation
energy (MACE) filter, the quadratic distance correlation classifier (QDCC), and the shift-invariant 2D
pattern-matching classifier. The basic structures of the SDF and the MACE filter are characterized in the
frequency domain by

(25)

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect5.html (1 of 3)22.08.2005 16:10:33

 Spatial Matched-Filter Classifiers for SAR

where H denotes the DFT of the spatial matched filter. The matrix X is composed of a set of target
training vectors obtained by taking the DFT of the target training images. The vector U represents a set
of constraints imposed on the values of the correlation peaks obtained when the training vectors are run
through the spatial matched filter. The matrix A represents a positive definite weighting matrix. A is an
identity matrix for SDF, and is the inverse of the following matrix D

(26)

where N is the number of training images and p is the dimension of the training vectors.

In the QDCC, the DFT of the spatial matched filter is expressed by

(27)

where m1 and m2 are means of the DFTs of the training images for classes 1 and 2, respectively. S is a
diagonal matrix defined by

(28)

where M1 and M2 are matrices with elements of m1 and m2 placed on the main diagonal and X and Y
i i

are ith training vectors from classes 1 and 2, respectively.

In the shift-invariant 2D pattern matching classifier, the correlation scores are calculated by

(29)

where T is the DFT of the decibel-normalized test image and R is the ith reference template.
i

Novak et al. (18) did extensive experiment with the high-resolution (1 ft×1 ft) fully polarimetric SAR
data. In the four-class classification experiment using four types of spatial matched filter classifiers, it is
reported that all targets are correctly classified (18).

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect5.html (2 of 3)22.08.2005 16:10:33

Spatial Matched-Filter Classifiers for SAR

About Wiley InterScience | About Wiley | Privacy | Terms & Conditions

Copyright© 1999-2005 John Wiley & Sons, Inc. All rights reserved.

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect5.html (3 of 3)22.08.2005 16:10:33

 Multisensor Fusion

Radar Target Recognition

Kie B. Eom, The George Washington University, Washington, DC

Encyclopedia of RF and Microwave Engineering

Copyright © 2005 John Wiley & Sons, Inc. All rights reserved.
DOI: 10.1002/0471654507.eme347
Article Online Posting Date: April 15, 2005

6. Multisensor Fusion
There has been research on multisensor fusion for target recognition (28). Some of the motivating
factors of such research are increased target illumination, increased coverage, and increased information
for recognition. Significant improvement in target recognition performance has been reported (13) when
multiple radar sources were utilized using sensor fusion approaches. Tenney and Sandell (30) developed
a theory for obtaining the distributed Bayesian decision rules. Chair and Varshney (5) presented an
optimal fusion structure given that detectors are independently designed. The target recognition using
multiple sensors is formulated as a two-stage decision problem in Ref. 13. A typical radar target
recognition approach using data fusion is illustrated in Fig. 8. After the prescreening, Single-source
classifications are performed first; then the fusion of decisions is performed.

Figure 8. A typical data fusion approach for target recognition (adapted from Hauter et al.
(13)).
[Full View]

The data fusion problem is treated as an m-hypothesis problem with individual source decisions being
the observations. The decision rule for m hypothesis is written as

(30)

For Bayes’ rule, g (u) is an a posteriori probability:

i

(31)

Since the a priori probability and the distribution of features cannot be estimated accurately, a heuristic
function is used (13). It is a direct extension of Bayesian approach introduced by Chair and Varshney
(5), and the function g (·) is generalized to include the full threshold range
i

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect6.html (1 of 2)22.08.2005 16:10:41

 Multisensor Fusion

(32)

where P0 and P1 are prior probabilities, W1 and W0 are the sets of all i such that and, {g (u)
i

>T } respectively, where T is the individual source threshold for partitioning decision regions, and the
i i

probabilities and are false-alarm rates and probabilities of detections of each local sensor. The
probabilities and are defined by the cumulative distribution function (cdf) for each decision
statistic. In practice, the cdf is quantized and estimated from training on the individual sensor's classifier
error probabilities. In a distributed scenario, the weighting can be computed at each sensor and
transmitted to the fusion center where they will be summed and compared to the decision threshold. In
Ref. 13, the data fusion approach is applied to multiple polarimetric channels of a SAR image, and
substantially improved classification performance is reported.

About Wiley InterScience | About Wiley | Privacy | Terms & Conditions

Copyright© 1999-2005 John Wiley & Sons, Inc. All rights reserved.

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect6.html (2 of 2)22.08.2005 16:10:41

 Summary

Radar Target Recognition

Kie B. Eom, The George Washington University, Washington, DC

Encyclopedia of RF and Microwave Engineering

Copyright © 2005 John Wiley & Sons, Inc. All rights reserved.
DOI: 10.1002/0471654507.eme347
Article Online Posting Date: April 15, 2005

7. Summary
In radar target recognition, different types of radar are employed for different applications. Each radar
system has different characteristics, and target recognition approaches using different radar are discussed.

About Wiley InterScience | About Wiley | Privacy | Terms & Conditions

Copyright© 1999-2005 John Wiley & Sons, Inc. All rights reserved.

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/sect7.html22.08.2005 16:10:42
 Bibliography

Radar Target Recognition

Kie B. Eom, The George Washington University, Washington, DC

Encyclopedia of RF and Microwave Engineering

Copyright © 2005 John Wiley & Sons, Inc. All rights reserved.
DOI: 10.1002/0471654507.eme347
Article Online Posting Date: April 15, 2005

Bibliography
1. J. S. Baras and S. I. Wolk, Model-based automatic target recognition from high-range-resolution
radar returns, SPIE Proc. 2234: 57–66 (1994). Links
2. M. Basseville, A. Benveniste, and A. S. Willsky, Multiscale autoregressive processes, Part I & II,
IEEE Trans. Signal Process. 40: 1915–1954 (1992). Links
3. B. Bhanu, Automatic target recognition: State of the art survey, IEEE Trans. Aerospace Electron.
Syst. 22(4): 364–379 (1986). Links
4. V. Cantoni, L. Lombardi, L. Cinque, S. Levialdi, and C. Guerra, Recognizing 2D objects by a multi-
resolution approach, Proc. Int. Conf. Pattern Recognition, Jerusalem, Israel, Oct. 1994, Vol. 3, pp.
310–316.
5. Z. Chair and P. K. Varshney, Optimal data fusion in multiple sensor detection systems, IEEE Trans.
Aerospace Electron. Syst. 22(1): 98–101 (1986). Links
6. N. C. Currie, R. D. Hayes, and R. N. Trebits, Millimeter-Wave Radar Clutter, Artech House, 1992.
7. I. Daubechies, The wavelet transform, time-frequency localization and signal analysis, IEEE Trans.
Inform. Theory. 36: 961–1005 (1990). Links
8. D. M. Dunn, W. H. Williams, and T. L. DeChaine, Aggregate versus subaggregate models in local
area forecasting, J. Am. Stat. Assoc. 71: 68–71 (1976). Links
9. R. Chellappa, Q. Zhang, C. Shekhar, P. Burlina, and K. Eom, On the positioning of multisensor
imagery for exploitation and target recognition, Proc. IEEE. 85(1): 120–138 (Jan. 1997). Links
10. K. B. Eom and R. Chellappa, Non-cooperative target classification using hierarchical modeling of
high range resolution radar signatures, IEEE Trans. Signal Process. 45(9): 2318–2327 (Sept. 1997).
Links
11. J. Geweke, Temporal aggregation in the multiple regression model, Econometrica. 46: 643–662
(1978). Links
12. C. W. J. Granger and M. J. Morris, Time series modeling and interpretation, J. Roy. Stat. Soc. A-
139:246–257 (1976). Links
13. A. Hauter, K. C. Chang, and S. Karp, Polarimetric fusion for synthetic aperture radar target
classification, Pattern Recogn. 30(5): 769–775 (1997). Links

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/bibliography.html (1 of 3)22.08.2005 16:10:44

Bibliography

14. S. Kingsley and S. Quegan, Understanding Radar Systems, McGraw-Hill, 1992.

15. D. C. McKee, C. Bandera, S. Ghosal, and P. J. Rauss, Model-based automatic target recognition
using hierarchical foveal machine vision, SPIE Proc. 2755: 70–79 (1996). Links
16. R. A. Mitchell and R. Dewall, Overview of high range resolution radar target identification, Proc.
Automatic Target Recognition Working Group Conf., Monterey, CA, Nov. 1994.
17. L. M. Novak, A comparison of 1-D and 2-D algorithms for radar target classification, Proc. IEEE
Int. Conf. Systems Engineering, Aug. 1991.
18. L. M. Novak, Radar target identification using spatial matched filters, Pattern Recogn. 27(4): 607–
617 (1994). Links
19. L. I. Perlovsky, W. H. Schoendorf, B. J. Burdick, and D. M. Tye, Model-based neural network for
target detection in SAR images, IEEE Trans. Image Process. 6(1): 203–216 (1997). Links
20. F. A. Pino, P. A. Morettin, and R. P. Mentz, Modelling and forecasting linear combinations of time
series, Int. Stat. Rev. 55: 295–313 (1987). Links
21. A. W. Rihaczek and S. J. Hershkowitz, Radar Resolution and Complex-Image Analysis, Artech
House, 1996.
22. O. Rioul, A discrete-time multiresolution theory, IEEE Trans. Acoust. Speech Signal Process. 41:
2591–2606 (1993). Links
23. M. W. Roth, Neural networks for extraction of weak targets in high clutter environments, IEEE
Trans. Syst. Man Cyber. 19(5): 1210–1217 (1989). Links
24. M. W. Roth, Survey of neural network technology for automatic target recognition, IEEE Trans.
Neural Networks. 1(1): 28–43 (1990). Links
25. M. A. Rubin, Application of fuzzy ARTMAP and ART-EMAP to automatic target recognition
using radar range profiles, Neural Networks. 8(7-8): 1109–1116 (1995). Links
26. M. L. Skolink, Introduction to Radar Systems, McGraw-Hill, 1980.
27. C. Stewart, Y. C. Lu, and V. Larson, A neural clustering approach for high resolution radar target
classification, Pattern Recogn. 27(4): 503–513 (1994). Links
28. N. S. Subotic, B. J. Thelen, J. D. Gorman, and M. F. Reiley, Multiresolution detection of coherent
radar targets, IEEE Trans. Image Process. 6(1): 21–35 (1997). Links
29. L. G. Telser, Discrete samples and moving sums in stationary stochastic processes, J. Am. Stat.
Assoc. 62: 484–499 (1967). Links
30. R. R. Tenney and N. R. Sandell, Detection with distributed sensors, IEEE Trans. Aerospace
Electron. Syst. 17: 501–510 (1981). Links
31. J. D. Wald, D. B. Krig, and T. DePersia, ATR: Problems and possibilities for the IU community,
Proc. ARPA Image Understanding Workshop, Jan. 1992, pp. 255–264.
32. D. R. Wehner, High Resolution Radar, Artech House, 1987.
33. W. Wei, The effect of temporal aggregation of parameter estimation in distributed lag model, J.
Econometr. 8: 237–246 (1978). Links

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/bibliography.html (2 of 3)22.08.2005 16:10:44

Bibliography

34. W. W. S. Wei and D. O. Stram, Disaggregation of time series models, J. Roy. Stat. Soc. B-52: 453–
467 (1990). Links
35. M. A. Wincek and G. C. Reinsel, An exact likelihood estimation procedure for regression ARMA
time series models with possibly non-consecutive data, J. Roy. Stat. Soc. B-48: 303–313 (1986).
Links
36. H. A. Zebker and R. M. Goldstein, Topographic mapping from interferometric synthetic aperture
radar observations, J. Geophys. Res. 91(B5): 4993–4999 (1986). Links

About Wiley InterScience | About Wiley | Privacy | Terms & Conditions

Copyright© 1999-2005 John Wiley & Sons, Inc. All rights reserved.

http://www.mrw.interscience.wiley.com/erfme/articles/eme347/bibliography.html (3 of 3)22.08.2005 16:10:44