Radar Signal Processing

http://www.mrw.interscience.wiley.com/erfme/articles/eme346/abstract.
html
Radar Signal Processing

Wei Ye, National University of Singapore
Tat Soon Yeo, National University of Singapore
Encyclopedia of RF and Microwave Engineering

Copyright © 2005 John Wiley & Sons, Inc. All rights reserved.
DOI: 10.1002/0471654507.eme346
Article Online Posting Date: April 15, 2005
Keywords: radar signal; signal processing; interference; signal-to-noise ratio; moving-target indicator;
high-resolution radar imaging
Abstract
Signal processing plays a key role in most modern radar operations. Radar signal processing commonly
refers to the techniques that are used to extract desired information about a target from the received
signals. Such information usually includes target presence and the position, velocity, trajectory, and
image of the target.
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/eme346/abstract.html22.08.2005 16:07:17
Fundamentals of Radar Signal Processing


DOI: 10.1002/0471654507.eme346
1. Fundamentals of Radar Signal Processing

The capability of radar in acquiring information about targets and environments has been greatly
improved over the decades, not only owing to the advent of sophisticated hardware, but also the
development of advanced signal processing techniques. Conventional radar is used mainly for detection,
localization, and tracking of targets. High-resolution imaging radar can provide more information such
as target size, shape, and image and hence can be used for target recognition, which is desired by
modern radar. Signal processing plays a key role in most modern radar operations. Radar signal
processing commonly refers to the techniques that are used to extract desired information about a target
from the received signals. Such information usually includes target presence and the position, velocity,
trajectory, and image of the target.
According to their objectives, the signal processing techniques can be divided into two categories. The
techniques in the first category are used mainly to enhance the useful signal and suppress interference,
thus enabling the radar to work satisfactorily in less-than-ideal environments. The techniques in the
second category are developed to improve the resolving capability of radar, thus enabling the radar to
extract complex information, such as target size and shape.
1.1. Signal Enhancement and Interference Suppression

Extraction of desired information from radar echoes is not an easy task under most circumstances, since
the wanted signal has to compete with unwanted signals, such as noise, clutter, and external interference.
Noise includes thermal noise and other noiselike disturbances and errors introduced into the receiving
channels of radar. It is normally modeled as a Gaussian random process with a uniform power spectrum
(white). Clutter is the echo from targets of no interest. For example, if the target of interest is an aircraft,
the echoes from the ground, clouds, and rain will produce clutter. Unlike that of noise, the spectrum of
clutter is not uniform and is determined by the Doppler frequency and strength of the clutter source.
External interference can be generated either by hostile jamming devices or other microwave equipment
operating nearby. Radars usually operate in such a combined environment. Different techniques are
required to suppress noise, clutter, and jamming since they have different properties.
We first consider coping with noise. In order to ensure reliable detection and processing, it is desirable
to maximize the signal-to-noise ratio (SNR). The optimum filter that achieves the maximum SNR is the
matched filter. The matched filter is widely used in radar systems and is often considered part of the
http://www.mrw.interscience.wiley.com/erfme/articles/eme346/sect1.html (1 of 5)22.08.2005 16:07:33

radar receiver, because the intermediate-frequency (IF) amplifier in the radar receiver is normally
designed as the matched filter. The matched filter is also an important element in pulse compression.
The theory of matched filter is discussed in detail in another article in this encyclopedia, Radar Signal
Detection. This article briefly introduces the concept of matched filter.
Denote the input signal and its Fourier transform by s(t) and S( w), respectively. The input noise is
assumed to be white and stationary. Its spectral density in watts/hertz is N0/2. The frequency response of
the matched filter is (1)
(1)
where the asterisk denotes the complex conjugate, k is a gain constant, and t0 is a time delay that makes
the filter physically realizable. The impulse response of the matched filter is the inverse Fourier
transform of Eq. (1), which is
(2)
So the impulse response is the image or time inverse of the input signal.
The SNR used here is defined as the ratio of the peak instantaneous signal power and the average noise
power. It is maximized at the output of the matched filter as SNR =2E/N0, where E is the signal
max
energy. It is an interesting fact that the output SNR is independent of the waveform shape of the input
signal.
Clutter reduction is also a major concern for many radar applications. Moving-target indication (MTI) is
an important technique that can discriminate moving targets from fixed or slowly moving clutter. MTI is
based on the phenomenon that the target moving with finite radial velocity produces a Doppler
frequency whereas the fixed target does not. So the difference in Doppler frequency can be utilized to
filter out the fixed clutter.
For radars operating in military environment, antijamming capability is a crucial performance indicator.
In this case, the technique of frequency-domain filtering is ineffective, since the interfering signal
usually has similar Doppler properties as the desired signal. However, a jamming source is confined to
within a certain spatial angle and is usually not in the same direction as the target. So it can be removed
by spatial filtering. The technique of spatial filtering is known as beamforming. By combining the
signals of individual elements of a phased-array antenna, it can synthesize a beam pattern whose
mainlobe points to the desired target and a very low sidelobe points to the jamming source.
However, no practical techniques can completely suppress these unwanted signals. The residual
interference will lead to a higher false-alarm rate when a fixed threshold is used to decide the presence
of a target. Having too many false alarms is unacceptable for the radar detector. Moreover, it also
increases the computational load of subsequent data processors. Therefore, a device is needed to control

the threshold adaptively in order to maintain an approximately constant false-alarm rate (CFAR).
1.2. Improvement of Resolving Capability

Radar has two main tasks: detection and information extraction. Early radar only has the capability to
extract simple target information, such as range, angle, and velocity. Modern radar is able to acquire
complicated information such as target size, shape, and image. This is largely a result of the
advancement of radar signal processing. Extracting such information requires the radar to have high
resolution in one or two (or even three) dimensions.
Range resolution is the ability of radar to distinguish adjacent targets in range direction. If the radar
transmits a short pulse of single frequency, range resolution is decided by the pulse duration T. In this
case, two scatterers adjacent in range can be distinguished by the radar if their echoes are separated in
time delay by more than the pulsewidth. So the range resolution is
(3)
where c is the speed of light and the factor 2 is introduced by the two-way time delay.
Thus, the shorter the pulse, the finer the range resolution. However, a very short pulse carries little
energy, and it is impractical, if not impossible, to generate short pulses of very large amplitudes. Hence,
a detection range would be shortened, due to fixed SNR requirement of the receiver. To overcome this
drawback, high-resolution radars usually adopt long-duration, wideband, coded pulses, which can be
compressed to very short pulses after reception by an operation called pulse compression.
High range resolution makes it possible for the radar to acquire information on target size and shape. On
the other hand, if the target image is needed, the radar must have high resolution in the azimuth direction
as well. For conventional radar, this corresponds to the angular resolution, which is actually the antenna
beamwidth, because the radar can distinguish two scatterers in the azimuth only if their angular
separation is greater than the antenna beamwidth. The beamwidth is decided by the ratio of the signal
wavelength to the aperture size of the antenna. The larger the aperture, the sharper the beam.
However, it is usually impractical to increase the aperture size of a real antenna to a great extent. For
instance, assuming a signal wavelength of 3 cm, a whopping 300 m antenna aperture is needed in order
to resolve two targets located 1 m apart at a range of 10 km. The synthetic aperture radar (SAR) is
developed to overcome this limitation. In SAR, the radar is installed on a moving platform, such as an
aircraft or a satellite. The radar transmits and receives a series of pulses while moving with the platform.
This is equivalent to obtaining discrete samples of a long aperture. The long antenna aperture can be
synthesized by a coherent summation of these samples after proper phase adjustment. Thus, SAR
achieves high resolution in azimuth in a practical way. It presents the image of the target by resolving
the backscattering intensity of each small range-azimuth cell.
1.3. Radar Signal Processor

In a radar system, the device that performs the operations of radar signal processing is termed the radar
signal processor. The radar signal processor is sometimes considered part of the radar receiver in early

days. In modern radars, it incorporates more and more important functions and is usually considered a
separate subsystem due to its significance.
Figure 1a shows the block diagram of the receiving and processing channel of the radar system. The
receiver accepts radiofrequency (RF) signals from the antenna and downconverts then to an intermediate
frequency. The signals are then amplified, converted to videofrequency (baseband), and sent to the
analog-to-digital converter (ADC). The digital samples of the signals are the input to the signal
processor. Finally, the output of the signal processor goes to the data processor and display devices.
Figure 1. (a) Position of the radar signal processor in the receiving and processing channel
of a radar system; (b) An inside view of the radar signal processor. The configurations vary
in practical processors.
[Full View]
Figure 1b gives an inside view of the radar signal processor, which encompasses all major functions of
radar signal processing. The configurations of practical signal processors may vary according to the
main function of radar. A typical signal processor usually does not contain all the components depicted
in the block diagram, and may not incorporate them in the given order.
1.4. Coherent Signal Processing

Most functions of the radar signal processor require coherent processing, which means that both signal
amplitudes and phases are utilized. Noncoherent processing only uses signal amplitudes. A device called
synchronous detector or phase-sensitive detector, as shown in Fig. 2, obtains coherent video signals. It
consists of two orthogonal channels denoted by I (in-phase) and Q (quadrature), which correspond to the
real and imaginary part of a signal, respectively.
Figure 2. The synchronous detector consists of two orthogonal channels, corresponding to

the real and imaginary parts of a signal. Its output is a complex video signal.
[Full View]
The input is usually an IF signal, which has the general form of
(4)
where A(t) is the amplitude or envelope, fc is the IF frequency, and f(t) is the phase. It is mixed with the
following orthogonal signals in the two channels:
(5)

The output of the multiplier in the I channel is
(6)
The first term has a doubled frequency and is thus removed by the subsequent lowpass filter (LPF). The
final output of the I channel is
(7)
Similarly, the output of the Q channel is
(8)
The complex video signal can be reconstructed as
(9)
This complex signal is the so-called complex envelope of s(t). It contains the information of the
amplitude as well as the phase.


Moving-Target Indication


DOI: 10.1002/0471654507.eme346
2. Moving-Target Indication
Discrimination of moving targets from strong, fixed clutter is necessary for many radar applications, for
example, air surveillance and air-traffic control. The MTI achieves this goal by means of the Doppler
effect generated by a moving target. The transmitted signal is a series of pulses with a pulse repetition
frequency (PRF) fr. The Doppler frequency fD of a moving target will introduce a varying phase term
into the received signal. The total phase of the received signal becomes
(10)
where f0 is the initial phase. The output video signals of the synchronous detector are
(11)
(12)
In the frequency domain, their spectra are centered at ±fD. On the other hand, a fixed target has zero
Doppler frequency. The different spectral locations makes it possible to separate moving targets from
fixed clutter.
The I or Q components or both can be used in the MTI process. If only one channel is used, the signal of
a moving target may be lost in some special cases. For example, a moving target generating a Doppler
frequency fD that equals half the sampling frequency may not be detected if the samples happened to fall
at the zero crossings. Using both I and Q channels can overcome this problem. In this case, the signals of
I and Q are fed into two separate MTI filters.
2.1. Delay-Line Cancelers

The delay-line canceler is a widely used form of MTI. Since the video signal of a fixed target does not
vary from pulse to pulse, it can be subtracted using adjacent pulses. The single canceler is the simplest

implementation of the MTI. Figure 3 shows its structure. The delay line therein introduces a delayed
time that exactly equals one interpulse period, which is the reciprocal of the PRF, that is, . The
single canceler utilizes only two pulses, and the output signal is
(13)
Figure 3. The single canceler utilizes only two pulses. The delay time Tr exactly equals
one interpulse period.
[Full View]
The frequency response of a single canceler can be obtained via the Fourier transform of Eq. (13)
(14)
where w=2 pfD, since the frequency of the video signal is actually its Doppler frequency, as shown in
Eqs. (11) and (12). So the frequency response of the single canceler is
(15)
The magnitude of the frequency response is
(16)
Figure 4 shows the filtering mechanism of the delay-line canceler. The solid line is the magnitude of the
frequency response of the single canceler. It forms notches at the zero Doppler frequency and the integer
multiples of fr. The moving target whose Doppler frequency is not zero will pass the filter.
Figure 4. The solid line is the frequency responses of the single canceler. It has rejection
notches at zero Doppler frequency and nfr, where n is an integer. The double canceler,
shown as the dashed line, provides broader notches and thus removes more clutter. Blind
speed occurs when the Doppler frequency of the target equals nfr.
[Full View]
The spectrum of completely fixed clutter is an impulse at zero Doppler frequency and can be removed
by the single canceler. However, practical ground clutter usually has a finite spectrum due to the slow

motion of clutter scatterers, for example, the trees swaying in the wind. The common model of the
ground clutter is the Gaussian-shaped power spectrum centered on zero Doppler frequency, as depicted
in Fig. 4. It can be seen that the notches of the single canceler are not wide enough to remove most of
the clutter spectrum. An improved solution is to use a double canceler.
A double canceler is formed by cascading two single cancelers and utilizes three signal pulses, as shown
in Fig. 5a. The frequency response of the double canceler is
(17)
The magnitude of the frequency response, depicted by the dashed line in Fig. 4, is
(18)
Compared with the single canceler, the double canceler provides broader rejection notch and thus
removes more clutter. The second equation of Eq. (17) indicates that the double canceler can be
implemented by a transversal filter, or finite impulse response (FIR) filter, as shown in Fig. 5b.
Figure 5. (a) Double canceler is constructed by cascading by two single cancelers. An

equivalent implementation is shown in (b), which is a transversal filter, or FIR filter.
[Full View]
It is also possible to construct a triple or multiple canceler using similar methods. One advantage of the
delay-line canceler is that it is able to process all target and clutter echoes, regardless of range.
2.2. Performance of Moving-Target Indicator

Many indicators have been used to describe MTI performance. Among them, the MTI improvement
factor is the most widely used. Other indicators used are clutter attenuation, MTI gain, subclutter
visibility, cancellation ratio, and clutter visibility ratio.
2.2.1. MTI Improvement Factor

This is defined (2) as “the signal-to-clutter ratio at the output (so/co) of the clutter filter divided by the
signal-to-clutter ratio at the input (si/ci) of the clutter filter, averaged uniformly over all target radial
velocities of interest.” The improvement factor accounts for both clutter attenuation and target gain. It
can be expressed mathematically as

(19)
For the clutter with a zero-mean Gaussian-shaped spectrum, the improvement factors of the single
canceler and the double canceler are given, respectively, as (3)
(20)
(21)
where sc is the standard deviation of the clutter power spectrum.
2.2.2. Clutter Attenuation

The ratio of clutter power at the input of the MTI to the power of residual clutter at the output.
2.2.3. MTI Gain

The ratio of signal power at the output of the MTI to signal power at the input, averaged over all target
radial velocities of interest.
From the definitions of the improvement factor, the clutter attenuation, and the MTI gain, the
relationship between them can be expressed by
(22)
where G is the MTI gain and CA is the clutter attenuation.
2.2.4. Subclutter Visibility

This is defined (2) as the ratio by which the target echo power may be weaker than the coincident clutter
echo power and still be detected with specified detection and false-alarm probabilities. Target and clutter
powers are measured on a single pulse return, and all target radial velocities are assumed equally likely.
Subclutter visibility measures the capability of the MTI to detect moving target in the presence of
clutter. For example, 20 dB of subclutter visibility means that the moving target can be detected even
though the clutter echo power is 100 times the target echo power.
2.3. Other Moving-Target Indicators

2.3.1. Blind Speed and Staggered-PRF MTI
As shown in Fig. 4, the frequency responses of the single and double canceler have periodic nulls at
multiples of PRF fr. If the moving target produces a Doppler frequency that exactly equals fr, or

multiples of fr, its echo will also be canceled by the MTI. In other words, radar is not able to detect a
target moving at one of these radial velocities. So such a speed is named a blind speed.
Blind speed is actually an inherent phenomenon of pulsed radars in which a sinusoidal signal with the
Doppler frequency fD is discretely sampled at the frequency fr. In the case that fD=nfr, the sampling
occurs at the same point in each corresponding Doppler cycle. Then the moving target looks as if it is
stationary.
The problem of blind speed can be alleviated by staggering the PRF, because a speed that is blind at one
PRF is generally not blind at another PRF. In this case, the interpulse period of the transmitted signal is
changed from pulse to pulse. Time delay in the MTI canceler is also changed correspondingly. This type
of MTI is called staggered-PRF MTI. It has a much higher first blind speed than that of the conventional
MTI.
In staggered-PRF MTI, the ratio of interpulse periods is usually expressed by a set of prime integers,
which has no common divisor other than 1. If N interpulse periods are staggered by d1: d2: ¼ :d , the
N
increase of the first blind speed V can be expressed by
(23)
where VB is the blind speed corresponding to the average interpulse periods.
2.3.2. Moving-Target Detector

The moving-target detector (MTD) is an enhanced version of traditional MTI. The basic structure of a
MTD is a MTI precanceler followed by a Doppler filterbank. The Doppler filterbank performs the
Doppler filtering or Doppler frequency analysis and is typically implemented by the fast Fourier
transform (FFT). A Doppler filterbank using N-pulse FFT divides the whole frequency band into N
parts, which are processed separately.
The Doppler filterbank provides the following advantages: (1) the SNR is improved since the signal in a
certain filter competes only with interference that can pass the filter, (2) the Doppler frequency of a
target can be measured to a high accuracy, (3) moving clutter with a nonzero mean of Doppler shift can
be rejected by separately adjusting the threshold of each filter, and (4) a clutter map can be generated
using the output from the filterbanks.
2.3.3. Adaptive MTI

The adaptive MTI is able to adaptively shift its rejection notch to the location of the clutter spectrum. It
is very useful for canceling moving clutter, whose Doppler frequency is not zero. An example of
adaptive MTI implementation is the time-averaged-clutter coherent airborne radar (TACCAR) (4). A
phase-error circuit estimates the phase change of clutter signals between pulses, which is caused by its
Doppler frequency. The averaged estimate of phase error is used to control the coherent oscillator

(COHO) to produce a phase-shifted reference signal for mixing with the return signal. This is equivalent
to shifting the Doppler center of the moving clutter to zero. Then a conventional MTI can be used to
cancel the clutter. Other implementation methods of adaptive MTI includes changing the weights of the
MTI filter and using modern spectral estimation method.
2.3.4. Airborne MTI

The airborne MTI (AMTI) refers to the MTI techniques utilized by a moving radar. The relative motion
between radar and fixed clutter displaces the Doppler spectrum of clutter form zero. Thus, the adaptive
MTI is one of the AMTI techniques. Another adverse effect of platform motion is the loss of signal
correlation between adjacent pulses, because the illuminated area is slightly displaced from one pulse to
the next. The effects of platform motion can be mitigated by physically or electrically displacing the
antenna phase center along the plane of the aperture. This technique is called displaced phase center
antenna (DPCA).
A relatively new technique is named spacetime adaptive processing (STAP). It simultaneously cancels
clutter in angle (space) and Doppler (time) domain. For example, the clutter coming from the antenna
mainlobe is spread out in Doppler frequency due to platform motion and is not easy to be canceled in the
Doppler domain. The STAP technique can effectively resolve the clutter within the beam width into
angle Doppler cells. Targets can be detected in cells that are not occupied by clutter.


Beam Forming


DOI: 10.1002/0471654507.eme346
3. Beam Forming
The radar often encounters jamming interference in the military environment. The jamming source
usually occupies the same frequency band as the desired signal and is difficult to eliminate by frequency
filtering. Fortunately, the spatial location of jamming source is usually different from that of the target.
Spatial filtering can thus be used to separate them in the space domain.
Spatial filtering in radar is achieved by forming a specific radar antenna beam pattern. A narrow
mainlobe means good selectivity, and low sidelobes means strong rejection capability. However,
producing an antenna with very low sidelobes in all directions is difficult to achieve in practice. High-
power interfering signals can still come into the receiver through sidelobes and degrade the desired
signal. Furthermore, the direction of jamming source may change with time. The beamforming
technique deals with this problem. It is able to adaptively form an antenna beam pattern with a very low
sidelobe in the direction of the interference.
Beam forming utilizes array antennas and considers only signal reception. The array elements are
discrete spatial samples over the entire antenna aperture. The signals received by each element are
weighted and summed to form the array output. When the weights of array elements can be adaptively
adjusted, it is called adaptive beamforming.
3.1. Linear Array Model and Classical Beamforming

An ideal linear array consists of identical, omnidirectional, equispaced sensors that are located along a
straight line. Figure 6 shows a linear array beamformer for processing narrowband signals. At time
instant k, each sensor samples the arriving signal. The output of the beam former is calculated as
(24)
where x (k) is the input signal of sensor n at time k. For the convenience of notation, a complex
n
conjugate is applied to the weight w . Equation (24) can be rewritten in the following vector form
n

Beam Forming
(25)
where w and x are N-dimensional column vectors and H denotes the transposed conjugate.
Figure 6. In the linear array model for processing narrowband signals, the signals received
by each element are weighted and summed to form the array output.
[Full View]
The array receives plane waves from a target if it is located in the far field. Assume that the signal has a
carrier frequency f and comes from direction q, which is defined as the angle of incidence with respect to
the broadside of the array. If two adjacent array elements are separated by a distance d, the signals they
received have a phase difference of
(26)
where c is the speed of light. In the following analysis, we assume a fixed carrier frequency f. So f( q, f)
is simply denoted by f( q). At time instant k, if the signal received by the first element is A(k)exp[j f1
(k)], that received by the nth element is A(k)exp{j[ f1(k)–(n–1) f( q)]}.
If the first element is used as reference, which means the phase of its signal, f1(k), is set to zero, then the
phase of each array element can be written as
(27)
This vector is termed the steering vector, which is also known as the direction vector, array manifold, or
array response vector. The output of the beamformer can be expressed as
(28)
The normalized output is denoted by
(29)
where r( q) is called the beamformer response. The square of r( q) is defined as the beam pattern.
In classical beamforming, the objective is to approximate the ideal response that is unity in a desired
direction and zero elsewhere. The classical beamformer, also known as the conventional beamformer, is

Beam Forming
used to receive a desired signal coming from a known direction q0. The solution is to choose the weight
as w=d( q0). From Eq. (29) the beamformer response is
(30)
where S (x) is defined as

N
(31)
and is called the periodic sinc function. The peak of the response is located at angle q0. So the mainlobe
of the beam points to q0. It is also said that the beam is steered to q0. In practice, amplitude tapering or
weighting is usually used to control the sidelobe levels. As r( q) is the summation of the phase-shifted
output of each element, the beamformer is also termed a phased array.
3.2. Adaptive Beamforming

3.2.1. Multiple Sidelobe Canceler
As stated earlier, strong jamming sources usually come into the receiver through antenna sidelobes. It is
thus desirable to form a beam pattern with nulls in the directions of the jamming sources. The multiple
sidelobe canceler (MSC) is the earliest beamformer that achieves this goal. An MSC consists of a main
antenna and one or more auxiliary antennas, as shown in Fig. 7. The mainlobe of the main antenna
points to the target of interest. Interfering signals are assumed to enter through its sidelobes. It is
assumed that the auxiliary antennas only receive interfering signals and noise. The weight of each
element can be adaptively adjusted to produce an output that estimates the interfering signals in the main
channel. Subtracting the output from the main channel by this estimate effectively cancels the
interference in the main channel.
Figure 7. The multiple sidelobe canceler consists of a main antenna and an auxiliary
adaptive beamformer. The latter estimates the interference in the main channel, which is
then subtracted from the main channel.
[Full View]
Denote the signal in the main channel by ym, the input of auxiliary antennas by xa, and the weight vector
by w. The output of the auxiliary array is ya=wHxa. Since ya is assumed not to contain the target signal,
successful cancellation of interference is equivalent to minimization the overall output power
:

Beam Forming
(32)
The weight that satisfies this equation is called the optimum weight, which is given by (5)
(33)
where is the covariance matrix of input data in auxiliary antennas, and is

the cross-correlation vector of xa and ym.
One should notice the assumption that the target signal is absent in the auxiliary antennas. This is
necessary. Otherwise, it will cause cancellation of target signal in the final output. However, the MSC
can still work well with slight signal loss if the target signal in the auxiliary antennas is very small.
3.2.2. Linearly Constrained Adaptive Beamforming

Another popular class of adaptive beamforming techniques is the linearly constrained adaptive
beamforming. It prevents the loss of desired signal by imposing linear constraints on the weights.
Equation (29) shows that the beamformer response to a signal from direction q is . If the
direction of interest is q0, the constraint can be expressed as , where f is constant. It ensures
that the desired signal from angle q0 will pass the beamformer with response f. Thus, minimizing the
overall output power, , will minimize the output produced by interference and noise
from directions other than q0. The linearly constrained beamforming can be expressed as
(34)
The solution of the optimum weight is
(35)
When f=1, the beamformer is also termed the minimum variance distortionless response (MVDR)
beamformer. There is only one constraint on w in Eq. (34). In fact, multiple constraints can also be used
to add more control on the beam former response (6).
The generalized sidelobe canceler (GSC) (7) is an alternative approach to the linearly constrained
adaptive beamforming. It converts the constrained optimization of Eq. (34) to unconstrained
optimization by separating out the constraint. Figure 8 shows the structure of the GSC. The upper or

Beam Forming
main branch is a classical beamformer, which has a fixed weight wc. The weight wc is chosen so that the
response of the main beamformer satisfies the constraint in Eq. (34). So the desired signal from angle q0
will pass the main beam former with desired response f. Interference and noise will produce some output
determined by wc. The lower branch is an adaptive beamformer, which is used to estimate the output
produced by interference and noise in the main beamformer. The estimate is then subtracted from the
main branch to cancel the interference and noise in the final output. This is similar to that in the MSC.
The difference is that there is a preprocessing matrix B in the lower branch of the GSC. The purpose of
B is to block the desired signal and thus to prevent signal cancellation in the final output. So B is named
the blocking matrix. For an array that is not presteered, B needs to satisfy . The columns of
B are linearly independent, and the number of columns L should be less than the number of array
elements.
Figure 8. The upper channel of the generalized sidelobe canceler is a classical

beamformer, which has a constant response to the desired signal. The lower channel is an
adaptive beamformer. It estimates the interference in the upper channel, which is then
subtracted from the upper channel. The blocking matrix B blocks desired signal and thus
prevent signal cancellation.
[Full View]
The overall output of the GSC is y=ym–ya, where ym and ya are the output of the upper and lower
beamformers, respectively. Since ya does not contain the desired signal, cancellation of interference and
noise is again equivalent to minimize the overall output power
(36)
The solution to the unconstrained optimization is
(37)
where is the covariance matrix of input data.
Besides the MSC and the GSC, there are some other criteria for constructing an adaptive beam former,
such as using a reference signal or maximizing the SNR.
3.2.3. Adaptive Algorithms

Calculation of the optimum weights for the adaptive beam formers discussed before requires knowledge
of second-order statistics, especially the covariance of the input signal. Such information is normally not
available and needs to be estimated from received data. The optimum weights are usually estimated and
adjusted by adaptive algorithms.

Beam Forming
It can be proved that the adaptive beamforming problem is equivalent to the adaptive filtering. So the
many adaptive algorithms developed for adaptive filtering can be utilized for adaptive beamforming.
One popular class of adaptive algorithms is based on the gradient, for example, the widely used least-
mean-square (LMS) algorithm. Another important class is based on the least-square estimation, such as
the sample matrix inversion (SMI) and the recursive least square (RLS) algorithm. There is a separate
article in this encyclopedia, Radar Signal Detection, that discusses the adaptive filtering technique. For
details of these adaptive algorithms, please refer to that article.


Constant False-Alarm Rate


DOI: 10.1002/0471654507.eme346
4. Constant False-Alarm Rate

A radar detector makes decision of target presence or absence from the echo of each resolution cell. If
the echo is stronger than a prespecified threshold, the decision of target presence is made. Otherwise, it
declares no target. The basic parameters of a radar detector are the probability of false alarm and the
probability of detection. The article Radar Signal Detection in this encyclopedia discusses basic theories
of radar detection. This article focuses on an advanced technique for the detector to maintain
approximately constant false-alarm rate (CFAR).
4.1. CFAR Principles

We first investigate the principles of CFAR in the background of Gaussian interference. In this case, the
interfering signals in both channels of I and Q, denoted by x and y respectively, are Gaussian processes
with zero mean and variance (power) of s2. This is a common model for noise and many types of
clutter. A linear envelope detector produces the signal envelope or magnitude, , which is
used for target detection. The r has a Rayleigh probability density function (pdf) as
(38)
The mean value of r is
(39)
The false-alarm probability PFA for a given threshold rt can be calculated as
(40)
The PFA is extremely sensitive to the changes of interference power s2. For example, doubling the

interference power can increase the value of PFA from 10–8 to 10–4. Hence, it is necessary to maintain
approximately constant false-alarm rate in order to ensure that the radar works well when the intensity of
interference changes. This can be achieved by adaptively adjusting the threshold according to the
interference power. An equivalent method is to normalize the interference signal so that its pdf will be
independent of its power s2.
Normalization of the Gaussian processes x and y is done by computing x/ s and y/ s. Correspondingly,

the normalized envelope is
(41)
Its pdf becomes
(42)
The false-alarm probability is
(43)
It can be seen that the new false-alarm probability is independent of the interference power s2. This is
the desired property of CFAR. Therefore, if s can be estimated from the received signals, the
normalization process r/ s will produce the constant false-alarm rate. From Eq. (39), we know that an
estimate of s can be obtained by estimating the mean value of r.
An alternative method is to use a square-law envelope detector. The magnitude squared of the envelope
is produced, denoted by q=r2, and used for deciding the presence of the target. The pdf of q can be
calculated from Eq. (38) as
(44)
The mean value of q is
(45)
Normalization of q can be done as

(46)
The pdf of the normalized variable qN is
(47)
In this case, the false-alarm probability for a preset threshold qt is
(48)
The false-alarm probability is independent of s2. So the CFAR property is achieved by the
normalization of q/ s2. Equation (45) indicates that an estimate of s2 can be obtained by estimating the
mean value of q.
The CFAR principle in the presence of Gaussian interference has been discussed for the linear and
square-law envelope detectors. For non-Gaussian interference, its envelope cannot be represented by the
Rayleigh pdf model. Specific models, such as the lognormal or the Weibull distribution must be used
according to the environment and application. In such cases, the CFAR can still be achieved by proper
normalization of the interference signal. The CFAR processing protects the radar detector from
excessive numbers of false alarms. The cost is that the probability of detection will be slightly reduced
compared with similar cases of thermal-noise-only environment.
4.2. Cell-Averaging CFAR

As analyzed before, for Gaussian interference, the normalization factor can be obtained by estimating
the mean value of the signal output from the envelope detector. One basic approach of CFAR is to
estimate the mean value by averaging the signals of a set of reference cells around the cell under test.
This technique is called cell-averaging CFAR (CA-CFAR). The reference cells can be either in range,
angle, or Doppler frequency. The basic assumption in CA-CFAR is that the interference is statistically
homogeneous over the reference cells and the cell under test.
Figure 9 shows the structure of the CA-CFAR processor that uses reference cells in range. The envelope
detector can have either a linear characteristic that produces the envelope r or a square-law characteristic
that produces the magnitude squared of envelope q. The received signals from the reference cells are
averaged to produce an estimate of their mean value. The immediate neighbors of the test cell are
sometimes discarded in case a strong target signal extended into them. The signal of the test cell is then
normalized by dividing it by the estimated mean and multiplying it with a scaling constant k. We know
from Eqs. (39) and (45) that for the linear detector and k=2 for the square-law detector.
Finally the normalized envelope of the test cell is compared with a preset threshold to decide if a target
is present.

Figure 9. A number of reference cells around the cell under test are used to estimate the
interference power in the cell-averaging CFAR processor. The estimate is used to
normalize the signal, which is equivalent to adaptively adjusting the threshold.
[Full View]
The CA-CFAR principle is applicable to both linear and square-law detection. The analysis that follows
is developed for square-law detection without loss of generality. Denote the magnitude squared of
envelope of the test cell by q0, and that of the reference cells by . These random
variables are statistically independent and have the same probability density function in a homogeneous
environment as given in Eq. (44). Their joint density function is
(49)
Denote the sum of the M reference cells by . The normalization of q0 is
(50)
q0N is then compared with a threshold qt, and the target-present decision is made when
(51)
In practice, only a limited number of reference cells M can be utilized to estimate the mean value.
Intuitively, the performance of CA-CFAR must be dependent on M. Substituting Eq. (50) into Eq. (51),
we get
(52)
For random variables having the density function given by Eq. (49), the ratio q0/q ,
M
called f variate, is known to have the following pdf (8):
(53)

The false-alarm probability can then be calculated as
(54)
It can be seen from Eq. (54) that the CFAR property can be achieved even for small values of M, since
the false-alarm probability is independent of the interference power s2. Equation (54) approximates the
ideal performance when M is very large such that
(55)
For a limited number of reference cells, higher signal-to-noise ratio is needed to achieve the required
false-alarm probability and detection probability. This is referred to as the CFAR loss.
For Swerling I and II targets, one can find the detection probability using similar methods. The signals
of a Swerling I or II target in I and Q channels are also modeled as a zero-mean Gaussian process with a
variance of . The overall variance is thus when the target is present. Denote the signal-to-
noise ratio by . A similar analysis shows the detection probability Pd is
(56)
and
(57)
For given PFA and Pd, the SNR for M reference cells, denoted by SNR , can be derived from Eqs. (54)
M
and (56). On the other hand, the SNR for M→∞, denoted by SNR , can be obtained from Eqs. (55) and
∞
(57). The CFAR loss is defined as the ratio of SNR and SNR . It is a function of PFA, Pd, and M. It
M ∞
has been shown (9) that for large M, the CFAR loss can be approximated by .
4.3. Ordered-Statistics CFAR

The CA-CFAR does not work well for nonhomogeneous interference. One example is when there are
interfering targets in the reference cells. Another case is when the background clutter varies dramatically
along either range or azimuth or both. Cell averaging can no longer produce the correct estimate, which
results in the loss of detection performance. The ordered-statistics CFAR (OS-CFAR) has been proven
to have better performance in such environments.
In the OS-CFAR, the threshold is determined by multiplying one of the ranked cell by a scaling factor.
Assume that the reference cells have the signals of after the envelope detector. They
are ranked as . The variable K is the rank of the cell to be used for
calculating the threshold, which is
(58)
The scaling factor a provides a mechanism to control the false-alarm probability.
4.4. Clutter-Map CFAR

The clutter-map CFAR is an effective approach for dealing with nonhomogeneous interference. It
utilizes the statistics of past observations on the test cell itself rather than that of the reference cells. It
assumes that the statistics of the test cell do not change during the observations so that time averaging
can be performed. In fact, the average of previous observations on each cell produces a clutter map—an
estimate of clutter power in each corresponding cell.
The principle of clutter-map CFAR is similar to that of the CA-CFAR. The basic analysis of CA-CFAR
is applicable to clutter-map CFAR. In clutter-map CFAR processing, the M reference samples are
obtained from M scans on the test cell. They are averaged in the same way as that in CA-CFAR to
produce an estimate of the interference power. Further processing is also similar to CA-CFAR. In a new
scan, if the target-present decision is made, the power estimate will not be updated. Otherwise, the new
sample is used to compute a new average.
A simplified implementation is to store only the estimate by previous samples, denoted by from
previous n–1 scans, and updates the estimate with the new sample q :
n
(59)
This is actually an exponential smoothing, since
(60)
The value of w determines the length of the exponential window. In practice, the length of the window
should be chosen to commensurate with the rate of change of the interference power.



Pulse Compression


DOI: 10.1002/0471654507.eme346
5. Pulse Compression
Pulse compression is a technique that obtains high range resolution while maintaining good detection
range. Signals used for pulse compression should have large bandwidth. There are two approaches to
increase the bandwidth: frequency modulation (FM) and phase coding. The device that performs the
pulse compression is the matched filter.
5.1. Frequency Modulation Waveforms

The linear FM signal is the earliest pulse compression waveform. Techniques for linear FM pulse
compression are well developed and widely used. Another relatively new technique adopts the stepped-
frequency waveforms.
5.1.1. Principles of Linear FM Pulse Compression

The linear FM signal, whose carrier frequency varies linearly during the pulse, is the most widely used
waveform. It is also known as the chirp signal and can be expressed by
(61)
where f0 is the carrier frequency. The constant m is called the chirp rate. The instantaneous frequency of
the signal is , which starts at and ends at . The signal bandwidth is B=
mT.
It can be shown (10) that the matched filter for the linear FM pulse compression can be expressed as
(62)
The h(t) has an inverse chirp rate as compared with the input signal. The output of the matched filter is

Pulse Compression
(63)
The envelope of the output signal is a sinc function:
(64)
For the sinc function, the peak–null pulsewidth is defined as the distance from the peak to the first null.
For sin x/x, the first null occurs at x= p. So the width of the compressed pulse is
(65)
It equals the reciprocal of the signal bandwidth. The pulse compression ratio is defined as the ratio of the
pulsewidth before and after the pulse compression processing. From Eq. (65), we have
(66)
Therefore, the pulse compression ratio is equal to the time–bandwidth product of the signal.
From Eq. (3), the range resolution after pulse compression is
(67)
It should be noted that the peak–null pulsewidth corresponds to the mainlobe width at –4 dB (power)
points. Another definition of pulsewidth is the mainlobe width at–3 dB, which is called half-power
pulsewidth and can be expressed as Kw/B, where Kw is a constant factor. For the sinc function,
Kw=0.866.

Pulse Compression
It is well known that the sinc function has high sidelobes. To reduce sidelobes, one approach is to apply
weighting in the compression filter. It can greatly reduce the sidelobes at the cost of mainlobe
broadening. Table 1 lists some popular weighting functions with the corresponding peak sidelobe levels
and mainlobe width at –3 dB. Another approach to reduce the sidelobes is to use nonlinear FM
waveforms. In this case, the spectrum of the transmitted signal is weighted, which also provides low
sidelobes in the compressed pulse.
Table 1. Properties of Weighting Functions
Mainlobe Width at –3 dB, Kw/

Weighting Function Peak Sidelobe Level (dB) B
Rectangular –13.26 0.886/B

Hanning –31.5 1.42/B
Hamming –42.5 1.32/B
Taylor, –35 1.19/B
Taylor, –40 1.25/B
Dolph–Chebyshev –40 1.20/B
5.1.2. Implementation of Linear FM Pulse Compression

The commonly used analog devices for linear FM pulse compression are the dispersive delay lines, in
which the signal delay is dependent on its frequency. The most popular one is the surface acoustic wave
(SAW) device. The electrical signal is converted into a surface wave at the input, which propagates
through the medium at acoustic speed. The dispersive property of the device achieves the matched
filtering. At the output the wave is converted back to the electrical signal. Another type of ultrasonic
device is the bulk acoustic wave device, which is less energy-efficient than the SAW device. These
ultrasonic devices can process signals with bandwidth greater than 1000 MHz (10).
Linear FM pulse compression can also be implemented by digital methods. The received signal is
sampled by an ADC and transformed into the frequency domain via FFT. The matched filtering is
performed in the frequency domain by multiplying the signal spectrum with the frequency response of
the matched filter. The result is then transformed back into the time domain.
A third technique is called stretch processing, which can process signals with very large time–bandwidth

Pulse Compression
products. Figure 10 shows its principle. Echoes from a set of scatterers are linear FM signals whose
starting points correspond to the ranges of each scatterer. They are mixed with a reference signal, which
is also a linear FM with the same chirp rate as that of the transmitted signal. The mixing (dechirping)
process removes the frequency modulation of the received signal. Each dechirped signal has a constant
frequency that corresponds to its range. An ADC then samples these dechirped signals. Finally, a
frequency analysis by FFT distinguishes each scatterer in the range. The stretch processing is also
widely used in high-resolution radars.
Figure 10. The stretch processing can reduce the signal bandwidth by dechirping the
return signal. The dechirped signal is then digitized and transformed by an FFT, which
accomplishes pulse compression.
[Full View]
5.1.3. Stepped-Frequency Waveforms

In this pulse compression technique, the radar transmits a burst of narrowband pulses, and the carrier
frequency changes discretely from pulse to pulse by a stepsize Df. Compared with the linear FM
waveform, the stepped-frequency waveforms can be regarded as discrete modulation in frequency. If
there are N pulses in a burst, the overall bandwidth of the transmitted signal is B=N Df.
Each received pulse is then downconverted to baseband, and a pair of I and Q samples is collected in a
desired range gate. The length of the range gate is determined by Df, which is . The
ensemble of I and Q samples represent results that the received signal within the desired range gate is
discretely sampled in the frequency domain over the bandwidth B. There is a condition to be satisfied.
That is, the duration of the baseband response of each pulse should extend the whole desired range gate.
This is ensured if , where Tp is the duration of each pulse.
Pulse compression is accomplished by applying the inverse discrete Fourier transform on the N pairs of I
and Q samples. The result is a high-resolution range profile with N complex samples over the desired
range gate. The length of the range profile is r, and the resolution is
(68)
Pulse compression using the stepped frequency avoids the requirements of wide instantaneous
bandwidth and high sampling rates. However, transmitting multiple pulses require relatively long time.
When there is large motion between the radar and the target, the motion-induced phase errors between
pulses must be compensated before the pulse compression can be performed.

Pulse Compression
5.2. Phase-Coded Waveforms

The phase-coded signal is a constant-amplitude sinusoid that is divided into N equal segments. The
phase of the sine wave in each segment is selected according to given code sequence. If the phase is set
at either 0 or p, it is called binary, or biphase, coding. If the code sequence contains more than two
values, it is called polyphase coding.
The pulse compression principle can be demonstrated by correlation processing, which is equivalent to
matched filtering. Consider a binary-coded signal with N segments. Denote the value of the kth segment
by a , which is either 1 or –1. The aperiodic autocorrelation function of this waveform can be written as
k
(69)
If m=0, then for each segment k, and the summation equals N. When m≠0, the value of the
autocorrelation function is much smaller than N. Thus the correlation processing produces a narrow
pulse, whose mainlobe locates at m=0 with the amplitude of N. The regions where m≠0 are sidelobes. A
well-designed code sequence can produce a pulse with very low sidelobes.
The Barker codes are the binary codes that have the lowest peak sidelobe. The aperiodic autocorrelation
of Barker codes has the value of either 0 or 1 for , and hence the peak sidelobe is 1. With respect
to the mainlobe, the peak sidelobe level is , or –22.3 dB. However, no Barker codes whose length is
greater than 13 have been found. The pulse compression ratio for Barker codes is limited to the
maximum value of 13. So it is important to find longer sequences with good sidelobe properties.
The minimum peak sidelobe codes are the binary sequences that attain the lowest peak sidelobes for a
given length. Of course, the Barker codes belong to this class of codes. Codes with length of 14–48 have
been found (11). For example, for length 48, the best peak sidelobe that can be achieved is 3, which is –
24.1 dB with respect to the mainlobe. Other types of long binary codes are random and pseudorandom
codes, in which the phase value of 0 or p are chosen in a random or essentially random manner, each
with a probability of 0.5.
All binary codes have the common weakness that they are quite sensitive to the Doppler shift in the
received signal. Their performance degrades significantly for large Doppler frequencies. The polyphase
codes exhibit better Doppler tolerance and have relatively good sidelobe characteristics. Frank codes and
P4 codes are commonly used polyphase codes. They can be thought of as discrete approximations to the
linear FM waveform.
The two devices to implement the phase-coded pulse compression are the all-range compressor and
cross correlator. The former is suitable for signals with varying delays, and works well for all target
range. The latter is usually used when the target range is approximately known.

Pulse Compression


Summary


DOI: 10.1002/0471654507.eme346
7. Summary
This article has discussed major techniques of radar signal processing. These techniques can be divided
into two categories. The first category includes the techniques for suppressing interference and
enhancing the useful signal. They ensure reliable detection and processing, and are listed as follows:
• Matched Filter
It is the optimum filter that achieves the maximum SNR.
• MTI
It discriminates moving target from fixed clutter by Doppler filtering.
• Beamforming
It mitigates against jamming sources by spatial filtering.
• CFAR
It maintains a constant false-alarm rate even if the clutter intensity varies.
The techniques in the second category are developed mainly for improving the resolving capability of
radar. They enable radar to acquire complex information like target size, shape, and image:
• Pulse Compression
It achieves high resolution in range direction without sacrificing the detection range.
• Radar Imaging
The synthetic aperture radar is developed to attain high resolution in azimuth direction. The signal
processing techniques in SAR are discussed in some details, including methods for image formation,
motion compensation, and autofocus.
It is foreseeable that radar will advance with the advancement of signal processing techniques and
hardware technology. It will be more robust when working in less-than-ideal environments and have
stronger capability for acquiring complex information.

Summary


High-Resolution Radar Imaging


DOI: 10.1002/0471654507.eme346
6. High-Resolution Radar Imaging

A radar image is the spatial distribution of microwave reflectivity corresponding to the object. Besides
high range resolution, an imaging radar also requires a high azimuth resolution. It thus requires a large
antenna aperture in azimuth direction. To avoid practical difficulties in constructing a large antenna, the
synthetic aperture technique is developed.
6.1. Principles of Synthetic Aperture Techniques

The azimuth resolution of a real antenna is discussed first. Consider the linear array model of Fig. 6. For
simplicity, we assumed that there is no beamsteering and uniform weighting is applied to each array
element, that is, . Similar to Eq. (30), the array response can be found as
(70)
where l=c/f is the wavelength. The r( q) is also periodic sinc function similar to Eq. (31). The peak of its
mainlobe is at q =0. Its first null occurs when sin q= l/(Nd), where Nd=D is the entire array aperture.
When D is very large, sin q is very small, and sin q≈ q. So the first null occurs at q = l/D. The peak–
null beamwidth, the distance from the peak to the first null, is thus
(71)
Although the result is derived from the array antenna, it is applicable to continuous-aperture antennas.
The azimuth resolution of a real antenna at range R is R l/D.
Synthesizing a large antenna aperture can be achieved by mounting a small-aperture antenna on a

moving platform. The antenna transmits and receives signals while the platform moves. The received
data can then be used to synthesize a large aperture. Such a system is called synthetic aperture radar
(SAR).

Figure 11 shows two common operating modes of SAR: strip map and spotlight. In stripmap SAR, the
antenna beam is fixed relative to the flight path. The moving antenna sweeps out a strip of terrain that is
parallel to the flight path. In spotlight SAR, the beam is steered to continuously illuminate a relatively
small patch of terrain. There is another widely used mode that is called inverse synthetic aperture radar
(ISAR). In ISAR, the radar is usually stationary (although it can also be moving), and the target being
imaged is moving. After motion compensation, the relative motion between the radar and the target is
similar to that of the spotlight SAR.
Figure 11. (a) Stripmap-mode SAR. Its resolution is limited by the antenna beamwidth b
q
due to the fixed antenna pointing. (b) Spotlight-mode SAR. The antenna tracks a small
imaging area. Its resolution can be very high, but imaging area is limited by antenna
beamwidth.
[Full View]
The azimuth resolution of SAR can be derived from Fig. 12 by finding the synthetic-array response. The
cross-range (azimuth) direction, denoted by x, is defined in the flight direction. SAR processing can be
of focused or unfocused type. In focused processing, the phases of the signals are compensated so that
each array element receives a signal from the origin O with the same phase. It is equivalent to
transforming the linear array into an arc array. The unfocused processing does not perform such phase
compensation and has a low resolution. The cross-range resolution is derived here only for focused
processing.
Figure 12. The cross-range resolution of SAR can be derived from this array model. It is
compensated to an arc array in the focused SAR processing.
[Full View]
The movement of the radar along the arc path introduces a relative rotation between the radar and the
target. The observing angle from radar to the origin O will change an amount of d q between two
adjacent elements. Assume that there are N elements in all, and the central element is the reference. The
nth element has an observing angle of .
Denote the distance from the origin to each element on the arc array by R0. Since only the cross-range
resolution is of interest, we consider a point P at (x, 0). It is in the far field with respect to the real
antenna and . So the distance from P to the nth element is

(72)
The phase of the signal caused by the two-way time delay is
(73)
The first term in Eq. (73) is a constant phase and can be removed. The term n d q is usually very small,
and thus n d q ≈n d q. So the array response is
(74)
It is a sinc function. The peak is at x=0, and the first null occurs at . So the cross-range
resolution measured by the peak–null beamwidth is
(75)
where is the total integration angle. The cross-range resolution expressed by Eq. (75) is
applicable to strip-map SAR, spotlight SAR, and ISAR. It can also be expressed by the synthetic
aperture length L since . Thus
(76)
In stripmap SAR, the synthetic aperture length L and the integration angle D q are limited by the
beamwidth of the real antenna b due to the fixed beamsteering. As shown in Fig. 11, the maximum
q
value of , which is expressed by Eq. (71). Therefore, the maximum cross-range resolution in
stripmap SAR is
(77)
where D is the aperture length of the real antenna.
6.2. Stripmap SAR Image Formation

The most common method for stripmap SAR image formation is the range–Doppler (RD) algorithm.
Consider the geometry of data collection shown in Fig. 13, where r denotes the slant range and x denotes
the cross-range. Assume that the transmitted pulse is
(78)
where a(t) is a rectangular window of unity amplitude over the pulse duration, and f(t) represents the
phase modulation. The distance from the radar to a target at x=0 varies according to
(79)
The received signal from this point after demodulation is
(80)
where s is the radar cross section of the target. The change of antenna gain is assumed to be small and
can be omitted. The normalized form of Eq. (80) is the impulse response:
(81)
where
(82)
which is the well-known range migration (RM). RM can be neglected if b is small. It is also omitted in
q
the unfocused processing. In the focused processing, the integration of the signal must be along the
curved locus, or the curved locus must be straightened (RM-compensated) before integration.
Figure 13. Data collection geometry of the stripmap SAR.

[Full View]

The return signal from an extended target s (x, r) is
(83)
where denotes a two-dimensional convolution. It can be shown (12) that Eq. (83) can be expressed by
(84)
where
(85)
(86)
where v is the along-track speed of the radar and d(x) is Dirac's delta function. It can be seen that the
received signal is obtained by sequentially convolving s(x, r) with two impulse response h1 and h2.
Therefore s(x, r) can be reconstructed by sequentially correlating the return signal sr(x, r) with h2 and h1.
Equation (86) shows that h2 is actually a one-dimensional (1D) function of range r, and has the same
waveform as the transmitted pulse. Correlation of sr(x, r) with h2 is thus the conventional pulse
compression in range.
The h1 is a 2D function, due to the RM effect. The RM compensation is usually implemented by an

interpolation in the range–Doppler domain, which straightens the curved locus. The signal is then
compressed in azimuth after the correction (12-14). To further reduce the phase errors, a technique
called secondary range compression (13-15) can be adopted.
Besides the RD algorithm, two wavenumber domain algorithms have also been developed for SAR
processing: the range migration algorithm (RMA) (16) and the chirp scaling algorithm (CSA) (17).
Denote the Fourier transform of h(t, x, r) by , where k and k are spatial frequencies and are
x r
usually called x and r wavenumbers; w is the temporal frequency, and .
The RMA first transforms the received data sr(x, t) to Sr(k , w) by a 2D Fourier transform. Then it maps
x
the data from the k – w domain to the k –k domain by a process called Stolt interpolation. A 2-D phase
x x r
compensation is performed in this domain, which corrects the range migration effect. The image is

finally constructed by a 2D inverse Fourier transform.
The RMA is capable of imaging a large terrain area with fine resolutions. Its drawback is that the Stolt
interpolation is time consuming. Figure 14 shows an example of SAR image processed by the RMA.
The raw data were collected by ESAR-2, and made available by German Aerospace Research
Establishment (DLR). The image resolution is about 3 m in range and 0.2 m in azimuth.
Figure 14. An example of an ESAR-2 stripmap SAR image. It is processed by the range-
migration algorithm. [The ESAR-2 data were made available by courtesy of the German
Aerospace Research Establishment (DLR).]
[Full View]
The difficulty of RM correction is that scatterers located at different positions have different migration
loci, and thus a large amount of computational effort is need. On the other hand, with its unique
procedures, the chirp scaling algorithm is able to adjust all the different loci to the same trajectory in the
2D spatial-frequency domain without much computational effort. Then the range-migration correction
can be easily performed by a phase multiplication to each point.
The CSA requires only FFTs an complex multiplies. No interpolation is needed. So it is computationally
efficient. However, it has been shown (18) that the CSA implements only the shift and linear
components of the Stolt mapping. So the RM compensation is not complete. Nonetheless, such
approximate compensation can produce satisfactory results for most applications.
6.3. Image Formation in Spotlight SAR and ISAR

The spotlight SAR and ISAR are developed mainly for imaging a small terrain or target with high
resolutions. To take the advantage of the small imaging area, spotlight SAR and ISAR systems usually
adopt the linear FM signal and the stretch processing technique described earlier. Stretch processing can
largely reduce the analog signal bandwidth, and thus requires low AD sampling rates, and small storage
memory.
Recall the principles of stretch processing depicted in Fig. 10. Here the reference signal for dechirping
(mixing) process is chosen to be the replica of the return signal from the center of the scene being
imaged. The return signals from the scene center after mixing have zero Doppler frequency, which
means that there is no relative radial motion between the radar and the scene center. So it is also referred
to as motion compensation with respect to the scene center. While this procedure compensates the
translational motion, it retains the rotational motion between the radar and the scene center, which forms
an ideal arc synthetic array.
It can be shown (18) that the motion compensation dechirpes received signals in both range and azimuth

directions and transforms the data into the k –k domain. As discussed earlier, this is the two-
r x
dimensional spatial-frequency (wave-number) domain. Thus, the image can be constructed by a 2D

inverse Fourier transform. The question is that the data samples in the k –k domain are not located on a
r x
rectangular grid. Rather, they are on the polar coordinates. The date support is a ribbon whose size is
determined by the transmitted bandwidth B and the integration angle D q, as shown in Fig. 15.
Figure 15. Support area of spotlight SAR signals after motion compensation to the scene
center. Its size is determined by the signal bandwidth and the integration angle. Data
samples are located on the polar format.
[Full View]
For a small scene size, if the values of B and D q are small, which means low resolutions in range and
azimuth, the small ribbon can be directly approximated by a rectangle. Then the 2D inverse Fourier
transform can be performed by two 1D inverse FFTs in range and azimuth. This method is referred to as
a rectangular format algorithm. A scatterer can be well focused by this algorithm if it does not move
through resolution cells during the whole observation. For large scene size and high resolutions, the
scatterer's motion through resolution cells will cause image defocus. So the data must be stored in the
polar format. Since the Fourier transform does not have fast algorithms in polar format, a 2D
interpolation is needed to map the data from polar to rectangular grid. Finally the image can be
constructed via two 1D inverse FFTs. This method is called polar format algorithm.
The range-migration algorithm and the chirp scaling algorithm discussed in the preceding section can be
utilized for spotlight SAR image formation. They do not require motion compensation to the scene
center. Another method known as convolution backprojection algorithm, which is widely used in
computer-aided tomography (CT), can also be used for spotlight SAR image formation, since the
geometry of data collection in spotlight SAR is very similar to that in CT (19).
6.4. Autofocus Techniques

Even in a well-design SAR system, it is still difficult to measure the relative motion between the radar
and the target with high accuracy over the whole synthetic aperture. The case is even worse in ISAR
because the target is usually non-cooperative. In such cases, phase errors will be introduced to the
received signals. Such phase errors are space-invariant, which means all the range bins have the same
phase error during one observation. Autofocus is a useful technique to estimate and remove the space-
invariant phase errors.

Many autofocus algorithm have been presented, such as the map drift (MD) algorithm (18), the
dominant scatterer algorithm (DSA) (20), the phase gradient autofocus (PGA) algorithm (21), and the
weighted least-square (WLS) algorithm (22).
Autofocus usually begins in the range-compressed phase history domain, where the data are compressed
in range but not compressed in azimuth. Suppose the strongest scatterer in the nth range bin has the
Doppler frequency f and the initial phase . Other smaller scatterers are treated as clutter. So the
n
phase of the signal at range bin n and aperture position m is
(87)
where f (m) is the phase fluctuation caused by clutter. Clutter is commonly assumed independent from
n
range bin to range bin. When corrupted by a phase error g(m), Eq. (87) becomes
(88)
where N is the number of range bins. g(m) is the same for all range bins, since it is space-invariant.
Autofocus is carried out in two steps: estimation and correction. The challenge is the estimation of g(m).
Once the estimate is obtained, correction can be done by multiplying the range-compressed phase
history data by exp .
The WLS algorithm is an optimal estimation of g(m) in the sense that it minimizes the variance of the
residual phase error. It is capable of estimating and removing all kinds of phase errors, no matter if they
are of low order, high order, or entirely random. The first step in the WLS algorithm is to shift the
strongest scatterer in each range bin to the center of the image, thus removing its Doppler frequency
offset. The initial phase y0, is a constant and can also be removed by proper processing. Then Eq. (88)
n
becomes
(89)
The task is to estimate the error g(m) from Eq. (89). Since f (m) causes estimation errors, the objective
n
of the WLS estimation is to minimize the variance of the estimation error. Rewriting Eq. (89) using
vector notations, we have
(90)

where , and
.
Denote the variance of the phase term f (m) by . The covariance matrix of f(m) is
n
(91)
The weighted least-square estimate of g(m) is
(92)
Figure 16 shows a real ISAR image of an aircraft that is auto-focused by the WLS algorithm. The image
resolution is approximately 0.4×0.4 m.
Figure 16. A real ISAR image of an aircraft. It is autofocused by the WLS algorithm.
[Full View]
The PGA is a widely used algorithm in SAR community. It has an iterative procedure for phase error
correction. The first step is the same as that in the WLS algorithm, that is, moving the strongest scatterer
of each range bin to the image center. The second step is the windowing of each shifted image. The
window is centered on the strongest scatterer. Windowing increases the SNR by preserving the width of
dominant blur while discarding scatterers that cannot contribute to the phase error estimation.
The next step is to estimate the phase error, or rather, the phase gradient or phase difference. Denote the
range-compressed phase history data by , where f (m) is the
n
phase fluctuation caused by the clutter within the window. A linear unbiased minimum variance estimate
of the gradient of the phase error is
(93)
Another estimate of the phase difference is
(94)

where Ð denotes the angle of the complex quantity computed on [– p, p]. This is a maximum likelihood
estimate.
The estimated phase gradient or phase difference is integrated to obtain the estimate of phase error
. It is used for correction after removing any bias and linear component. Then the PGA goes back
to its first step and adopts a narrower window. The process is repeated until the image is well focused.


Bibliography


DOI: 10.1002/0471654507.eme346
Bibliography
1. M. I. Skolnik, Introduction to Radar Systems, 2nd ed., McGraw-Hill, New York, 1980.
2. J. Radatz, The IEEE Standard Dictionary of Electrical and Electronics Terms, 6th ed., IEEE, New
York, 1997.
3. D. K. Barton and S. A. Leonov, Radar Technology Encyclopedia, Artech House, Boston, 1997.
4. M. I. Skolnik, Radar Handbook, 2nd ed., McGraw-Hill, New York, 1990.
5. S. Haykin and A. Steinhardt, Adaptive Radar Detection and Estimation, Wiley, New York, 1992.
6. B. D. Van Veen and K. M. Buckley, Beamforming: A versatile approach to spatial filtering, IEEE
Acoust. Speech Signal Process. Mag. 5(2): 4–24 (1988). Links
7. L. J. Griffiths and C. W. Jim, An alternative approach to linearly constrained adaptive
beamforming, IEEE Trans. Anten. Propag. 30: 27–34 (1982). Links
8. R. Nitzberg, Adaptive Signal Processing for Radar, Artech House, Boston, 1992.
9. N. Levanon, Radar Principles, Wiley, New York, 1988.
10. D. R. Wehner, High-Resolution Radar, 2nd ed., Artech House, Boston, 1995.
11. F. E. Nathanson, Radar Design Principles, 2nd ed., McGraw-Hill, New York, 1990.
12. C. Wu, K. Y. Liu, and M. Y. Jin, Modeling and a correlation algorithm for spaceborne SAR signals,
IEEE Trans. Aerospace Electron. Syst. 18: 563–575 (1982). Links
13. M. Y. Jin and C. Wu, A SAR correlation algorithm which accommodates large range migration,
IEEE Trans. Geosci. Remote Sens. 22: 592–597 (1984). Links
14. R. Bamler, A comparison of range-doppler and wavenumber domain SAR focusing algorithms,
IEEE Trans. Geosci. Remote Sens. 30: 706–713 (1992). Links
15. F. H. Wong and I. G. Cumming, Error sensitivities of a secondary range compression algorithm for
processing squinted satellite SAR data, Proc. IEEE Geosci. Remote Sens. Symp. (IGARSS’89),
Vancouver, 1989, pp. 1702–1706.
16. C. Cafforio, C. Prati, and F. Rocca, SAR data focusing using seismic migration techniques, IEEE
Trans. Aerospace Electron. Syst. 27: 194–207 (1991). Links
http://www.mrw.interscience.wiley.com/erfme/articles/eme346/bibliography.html (1 of 2)22.08.2005 16:09:35

Bibliography
17. R. K. Raney et al., Precision SAR processing using chirp scaling, IEEE Trans. Geosci. Remote
Sens. 32: 786–799 (1994). Links
18. W. G. Carrara, R. S. Goodman, and R. M. Majewski, Spotlight Synthetic Aperture Radar: Signal
Processing Algorithms, Artech House, Boston, 1995.
19. C. V. Jakowatz, Jr. et al., Spotlight-Mode Synthetic Aperture Radar: A Signal Processing
Approach, Kluwer Academic, Boston, 1996.
20. C. C. Chen and H. C. Andrews, Target-motion-induced radar imaging, IEEE Trans. Aerospace
Electron. Syst. 16: 2–14 (1980). Links
21. D. E. Wahl et al., Phase gradient autofocus—a robust tool for high resolution SAR phase
correction, IEEE Trans. Aerospace Electron. Syst. 30: 827–835 (1994). Links
22. W. Ye, T. S. Yeo, and Z. Bao, Weighted least square estimation of phase errors for SAR/ISAR
autofocus, IEEE Trans. Geosci. Remote Sens. 37: 2487–2494 (1999). Links

http://www.mrw.interscience.wiley.com/erfme/articles/eme346/bibliography.html (2 of 2)22.08.2005 16:09:35

Radar Signal Processing

Uploaded by

Copyright:

Available Formats

Radar Signal Processing

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Radar Signal Processing

Uploaded by

Copyright:

Available Formats

http://www.mrw.interscience.wiley.com/erfme/articles/eme346/abstract.